Add remote code for Diffusers

README.md

@@ -105,27 +105,11 @@ Go to [ComfyUI-Animemory-Loader](https://github.com/animEEEmpire/ComfyUI-Animemo
 
 3.Diffusers inference.
 
-- The pipeline has not been merged yet. Please use the following code to setup the environment.
-```shell
-git clone https://github.com/huggingface/diffusers.git
-git clone https://github.com/animEEEmpire/diffusers_animemory
-cp diffusers_animemory/* diffusers -r
-
-# Method 1: Re-install diffusers. (Recommended)
-cd diffusers
-pip install .
-
-# Method 2: Call it locally. Change `YOUR_PATH` to the directory where you just cloned `diffusers` and `diffusers_animemory`.
-import sys
-sys.path.insert(0, 'YOUR_PATH/diffusers/src')
-```
-- And then, you can use the following code to generate images.
-
 ```python
-from diffusers import AniMemoryPipeline
+from diffusers import DiffusionPipeline
 import torch
 
-pipe = AniMemoryPipeline.from_pretrained("animEEEmpire/AniMemory-alpha", torch_dtype=torch.bfloat16)
+pipe = DiffusionPipeline.from_pretrained("animEEEmpire/AniMemory-alpha", trust_remote_code=True, torch_dtype=torch.bfloat16)
 pipe.to("cuda")
 
 prompt = "一只凶恶的狼，猩红的眼神，在午夜咆哮，月光皎洁"

model_index.json

@@ -1,5 +1,5 @@
 {
-  "_class_name": "AniMemoryPipeLine",
+  "_class_name": ["pipeline_animemory", "AniMemoryPipeline"],
   "_diffusers_version": "0.32.0.dev0",
   "feature_extractor": [
     null,
@@ -11,15 +11,15 @@
     null
   ],
   "scheduler": [
-    "diffusers",
+    "scheduling_euler_ancestral_discrete_x_pred",
     "EulerAncestralDiscreteXPredScheduler"
   ],
   "text_encoder": [
-    "animemory",
+    "animemory_t5",
     "AniMemoryT5"
   ],
   "text_encoder_2": [
-    "animemory",
+    "animemory_altclip",
     "AniMemoryAltCLip"
   ],
   "tokenizer": [
@@ -35,7 +35,7 @@
     "UNet2DConditionModel"
   ],
   "vae": [
-    "animemory",
+    "modeling_movq",
     "MoVQ"
   ]
 }

pipeline_animemory.py

@@ -0,0 +1,1771 @@
+# Copyright 2024 AniMemory Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+import numpy as np
+import PIL.Image
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import (
+    CLIPImageProcessor,
+    CLIPVisionModelWithProjection,
+    XLMRobertaTokenizerFast,
+)
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import (
+    FromSingleFileMixin,
+    IPAdapterMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+)
+from diffusers.models import ImageProjection, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    FusedAttnProcessor2_0,
+    XFormersAttnProcessor,
+)
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+
+from diffusers.utils import BaseOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+@dataclass
+class AniMemoryPipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+# TODO: update prompt case
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import AniMemoryPipeline
+
+        >>> pipe = AniMemoryPipeline.from_pretrained("animEEEmpire/AniMemory-alpha", torch_dtype=torch.bfloat16)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "一只凶恶的狼，猩红的眼神，在午夜咆哮，月光皎洁"
+        >>> negative_prompt = "nsfw, worst quality, low quality, normal quality, low resolution, monochrome, blurry, wrong, Mutated hands and fingers, text, ugly faces, twisted, jpeg artifacts, watermark, low contrast, realistic"
+        >>> image = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     num_inference_steps=40,
+        ...     height=1024,
+        ...     width=1024,
+        ...     guidance_scale=6.0,
+        ... ).images[0]
+        >>> image.save("output.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    r"""
+    Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
+    Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+
+    Args:
+        noise_cfg (`torch.Tensor`):
+            The predicted noise tensor for the guided diffusion process.
+        noise_pred_text (`torch.Tensor`):
+            The predicted noise tensor for the text-guided diffusion process.
+        guidance_rescale (`float`, *optional*, defaults to 0.0):
+            A rescale factor applied to the noise predictions.
+
+    Returns:
+        noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
+    """
+    std_text = noise_pred_text.std(
+        dim=list(range(1, noise_pred_text.ndim)), keepdim=True
+    )
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = (
+        guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    )
+    return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError(
+            "Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values"
+        )
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(
+            inspect.signature(scheduler.set_timesteps).parameters.keys()
+        )
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(
+            inspect.signature(scheduler.set_timesteps).parameters.keys()
+        )
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+def split_input_ids(
+    input_ids,
+    attention_mask,
+    start,
+    model_max_length,
+    bos_token_id,
+    eos_token_id,
+    pad_token_id,
+):
+    iids_list = []
+    mask_list = []
+    if start > 0:
+        cur_input_ids = input_ids[start - 1 :]
+        cur_input_ids[0] = bos_token_id
+        if attention_mask is not None:
+            cur_attention_mask = attention_mask[start - 1 :]
+            cur_attention_mask[0] = 1
+    else:
+        cur_input_ids = input_ids
+        if attention_mask is not None:
+            cur_attention_mask = attention_mask
+    n = len(cur_input_ids)
+
+    for i in range(1, n - model_max_length + 2, model_max_length - 2):
+        ids_chunk = (
+            cur_input_ids[0].unsqueeze(0),
+            cur_input_ids[i : i + model_max_length - 2],
+            cur_input_ids[-1].unsqueeze(0),
+        )
+        ids_chunk = torch.cat(ids_chunk)
+        if attention_mask is not None:
+            mask_chunk = (
+                cur_attention_mask[0].unsqueeze(0),
+                cur_attention_mask[i : i + model_max_length - 2],
+                cur_attention_mask[-1].unsqueeze(0),
+            )
+            mask_chunk = torch.cat(mask_chunk)
+
+        if ids_chunk[-2] != eos_token_id and ids_chunk[-2] != pad_token_id:
+            ids_chunk[-1] = eos_token_id
+            if attention_mask is not None:
+                mask_chunk[-1] = 1
+        if ids_chunk[1] == pad_token_id:
+            ids_chunk[1] = eos_token_id
+            if attention_mask is not None:
+                mask_chunk[1] = 1
+
+        iids_list.append(ids_chunk)
+        if attention_mask is not None:
+            mask_list.append(mask_chunk)
+
+    return iids_list, mask_list if len(mask_list) > 0 else None
+
+
+# Modified from [library.train_util.get_input_ids](https://github.com/kohya-ss/sd-scripts/blob/e5ac09574928ec02fba5fe78267764d26bb7faa6/library/train_util.py#L795)
+def get_input_ids(
+    caption,
+    tokenizer,
+    tokenizer_max_length,
+    dense_caption_split_method,
+    chunk,
+    punctuation_ids,
+):
+    prompt_tokens = tokenizer(
+        caption,
+        max_length=tokenizer_max_length,
+        padding="max_length",
+        truncation=True,
+        return_tensors="pt",
+    )
+    input_ids = prompt_tokens["input_ids"].squeeze(0)
+    attention_mask = prompt_tokens["attention_mask"].squeeze(0)
+
+    if not chunk:
+        return input_ids[None, ...], attention_mask[None, ...]
+
+    iids_list = []
+    mask_list = []
+
+    if dense_caption_split_method == "length_split":
+        iids_list, mask_list = split_input_ids(
+            input_ids,
+            attention_mask,
+            0,
+            tokenizer.model_max_length,
+            tokenizer.bos_token_id,
+            tokenizer.eos_token_id,
+            tokenizer.pad_token_id,
+        )
+    elif dense_caption_split_method == "punctuation_split":
+        can_split_tensor = torch.zeros_like(input_ids)
+        for punctuation_id in punctuation_ids:
+            can_split_tensor = torch.logical_or(
+                can_split_tensor, input_ids == punctuation_id
+            )
+        can_split_index = (
+            [0]
+            + [i[0] for i in torch.nonzero(can_split_tensor).tolist()]
+            + [len(input_ids) - 1]
+        )
+        start = 1
+        end = 1
+
+        new_can_split_index = []
+        for i in range(len(can_split_index) - 1):
+            pre = can_split_index[i]
+            new_can_split_index.append(pre)
+            nxt = can_split_index[i + 1]
+            cur = pre + tokenizer.model_max_length - 2
+            while cur < nxt:
+                new_can_split_index.append(cur)
+                cur = cur + tokenizer.model_max_length - 2
+        new_can_split_index.append(can_split_index[-1])
+        can_split_index = new_can_split_index
+
+        for i in can_split_index:
+            if i - start + 1 > tokenizer.model_max_length - 2:
+                if end == start:
+                    end = start + (tokenizer.model_max_length - 2)
+                ids_chunk = torch.tensor(
+                    [tokenizer.pad_token_id] * tokenizer.model_max_length,
+                    dtype=torch.int64,
+                )
+                ids_chunk[0] = tokenizer.bos_token_id
+                ids_chunk[1 : 1 + end - start] = input_ids[start:end]
+                ids_chunk[1 + end - start] = input_ids[-1]
+                mask_chunk = torch.zeros(tokenizer.model_max_length).to(torch.int64)
+                mask_chunk[0] = 1
+                mask_chunk[1 : 1 + end - start] = attention_mask[start:end]
+                mask_chunk[1 + end - start] = attention_mask[-1]
+                if ids_chunk[1] == tokenizer.pad_token_id:
+                    ids_chunk[1] = tokenizer.eos_token_id
+                    mask_chunk[1] = 1
+                if tokenizer.eos_token_id not in ids_chunk:
+                    ids_chunk[1 + end - start] = tokenizer.eos_token_id
+                    mask_chunk[1 + end - start] = 1
+                iids_list.append(ids_chunk)
+                mask_list.append(mask_chunk)
+                if len(iids_list) == 3:
+                    break
+                start = end
+            end = i + 1
+
+        if len(iids_list) == 0:
+            iids_list, mask_list = split_input_ids(
+                input_ids,
+                attention_mask,
+                0,
+                tokenizer.model_max_length,
+                tokenizer.bos_token_id,
+                tokenizer.eos_token_id,
+                tokenizer.pad_token_id,
+            )
+        elif len(iids_list) == 1:
+            iids_list1, mask_list1 = split_input_ids(
+                input_ids,
+                attention_mask,
+                start,
+                tokenizer.model_max_length,
+                tokenizer.bos_token_id,
+                tokenizer.eos_token_id,
+                tokenizer.pad_token_id,
+            )
+            iids_list = (iids_list + iids_list1)[:3]
+            mask_list = (mask_list + mask_list1)[:3]
+        elif len(iids_list) == 2:
+            iids_list1, mask_list1 = split_input_ids(
+                input_ids,
+                attention_mask,
+                start,
+                tokenizer.model_max_length,
+                tokenizer.bos_token_id,
+                tokenizer.eos_token_id,
+                tokenizer.pad_token_id,
+            )
+            iids_list = (iids_list + iids_list1)[:3]
+            mask_list = (mask_list + mask_list1)[:3]
+    else:
+        raise NotImplementedError
+
+    input_ids = torch.stack(iids_list)
+    attention_mask = torch.stack(mask_list)
+
+    return input_ids, attention_mask
+
+
+class AniMemoryPipeline(
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    FromSingleFileMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+    IPAdapterMixin,
+):
+    # TODO: review
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+    Args:
+        vae ([`MoVQ`]):
+            Variational Auto-Encoder (VAE) Model. AniMemory uses
+            [MoVQ](https://github.com/ai-forever/Kandinsky-3/blob/main/kandinsky3/movq.py)
+        text_encoder ([`AniMemoryT5`]):
+            Frozen text-encoder. AniMemory builds based on
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel).
+        text_encoder_2 ([`AniMemoryAltCLip`]):
+            Second frozen text-encoder. AniMemory builds based on
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection).
+        tokenizer (`XLMRobertaTokenizerFast`):
+            Tokenizer of class
+            [XLMRobertaTokenizerFast](https://huggingface.co/docs/transformers/v4.46.3/en/model_doc/xlm-roberta#transformers.XLMRobertaTokenizerFast).
+        tokenizer_2 (`XLMRobertaTokenizerFast`):
+            Second Tokenizer of class
+            [XLMRobertaTokenizerFast](https://huggingface.co/docs/transformers/v4.46.3/en/model_doc/xlm-roberta#transformers.XLMRobertaTokenizerFast).
+        unet ([`UNet2DConditionModel`]):
+            Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`EulerAncestralDiscreteXPredScheduler`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings shall be forced to always be set to 0.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+    _optional_components = [
+        "tokenizer",
+        "tokenizer_2",
+        "text_encoder",
+        "text_encoder_2",
+        "image_encoder",
+        "feature_extractor",
+    ]
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+        "negative_pooled_prompt_embeds",
+        "negative_add_time_ids",
+    ]
+
+    def __init__(
+        self,
+        vae: "MoVQ",  # type: ignore
+        text_encoder: "AniMemoryT5",  # type: ignore
+        text_encoder_2: "AniMemoryAltCLip",  # type: ignore
+        tokenizer: XLMRobertaTokenizerFast,
+        tokenizer_2: XLMRobertaTokenizerFast,
+        unet: UNet2DConditionModel,
+        scheduler: "EulerAncestralDiscreteXPredScheduler",  # type: ignore
+        image_encoder: CLIPVisionModelWithProjection = None,
+        feature_extractor: CLIPImageProcessor = None,
+        force_zeros_for_empty_prompt: bool = True,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.register_to_config(
+            force_zeros_for_empty_prompt=force_zeros_for_empty_prompt
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.default_sample_size = self.unet.config.sample_size
+
+        self.unet.time_proj.downscale_freq_shift = 1
+
+        self.scheduler.config.clip_sample = False
+        self.scheduler.config.timestep_spacing = "linspace"
+        self.scheduler.config.prediction_type = "sample"
+        self.scheduler.rescale_betas_zero_snr()
+
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        if device is None:
+            device = self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(
+            self, StableDiffusionXLLoraLoaderMixin
+        ):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder, lora_scale)
+
+            if self.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = (
+            [self.tokenizer, self.tokenizer_2]
+            if self.tokenizer is not None
+            else [self.tokenizer_2]
+        )
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2]
+            if self.text_encoder is not None
+            else [self.text_encoder_2]
+        )
+
+        punctuation_ids = [
+            [5, 4, 74, 32, 38, 4730, 30, 4, 74, 32, 38, 4730],
+            [5, 4, 74, 32, 38, 4730, 30, 4, 74, 32, 38, 4730],
+        ]
+        max_token_length = 227
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: process multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            text_encoder_idx = 0
+            for prompt, tokenizer, text_encoder in zip(
+                prompts, tokenizers, text_encoders
+            ):
+                text_input_ids, attention_mask = get_input_ids(
+                    prompt,
+                    tokenizers[text_encoder_idx],
+                    max_token_length,
+                    "punctuation_split",
+                    False if text_encoder_idx == 0 else True,
+                    punctuation_ids[text_encoder_idx],
+                )
+
+                tk_len = text_input_ids.shape[-1]
+                text_input_ids = text_input_ids.reshape((-1, tk_len))
+                attention_mask = attention_mask.reshape((-1, tk_len))
+
+                prompt_embeds, pooled_output = text_encoder(
+                    text_input_ids.to(device), attention_mask.to(device)
+                )
+
+                if text_encoder_idx == 1:
+                    tmp_ids = text_input_ids.reshape(-1, 3, text_input_ids.shape[-1])
+                    _, n2, tk_len2 = tmp_ids.size()
+                    prompt_embeds = prompt_embeds.reshape(
+                        (-1, n2 * tk_len2, prompt_embeds.shape[-1])
+                    )
+                    if n2 > 1:
+                        states_list = [prompt_embeds[:, 0].unsqueeze(1)]
+                        for i in range(
+                            1,
+                            max_token_length,
+                            tokenizers[text_encoder_idx].model_max_length,
+                        ):
+                            states_list.append(
+                                prompt_embeds[
+                                    :,
+                                    i : i
+                                    + tokenizers[text_encoder_idx].model_max_length
+                                    - 2,
+                                ]
+                            )
+                        states_list.append(prompt_embeds[:, -1].unsqueeze(1))
+                        prompt_embeds = torch.cat(states_list, dim=1)
+
+                        pooled_prompt_embeds = pooled_output[::n2]
+
+                prompt_embeds_list.append(prompt_embeds)
+                text_encoder_idx += 1
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = (
+            negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        )
+        if (
+            do_classifier_free_guidance
+            and negative_prompt_embeds is None
+            and zero_out_negative_prompt
+        ):
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            negative_prompt = (
+                batch_size * [negative_prompt]
+                if isinstance(negative_prompt, str)
+                else negative_prompt
+            )
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2]
+                if isinstance(negative_prompt_2, str)
+                else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            text_encoder_idx = 0
+            for negative_prompt, tokenizer, text_encoder in zip(
+                uncond_tokens, tokenizers, text_encoders
+            ):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(
+                        negative_prompt, tokenizer
+                    )
+
+                negative_text_input_ids, negative_attention_mask = get_input_ids(
+                    negative_prompt,
+                    tokenizers[text_encoder_idx],
+                    max_token_length,
+                    "punctuation_split",
+                    False if text_encoder_idx == 0 else True,
+                    punctuation_ids[text_encoder_idx],
+                )
+
+                tk_len = negative_text_input_ids.shape[-1]
+                negative_text_input_ids = negative_text_input_ids.reshape((-1, tk_len))
+                negative_attention_mask = negative_attention_mask.reshape((-1, tk_len))
+
+                negative_prompt_embeds, negative_pooled_ouput = text_encoder(
+                    negative_text_input_ids.to(device),
+                    negative_attention_mask.to(device),
+                )
+
+                if text_encoder_idx == 1:
+                    negative_tmp_ids = negative_text_input_ids.reshape(
+                        -1, 3, negative_text_input_ids.shape[-1]
+                    )
+                    _, n2, tk_len2 = negative_tmp_ids.size()
+                    negative_prompt_embeds = negative_prompt_embeds.reshape(
+                        (-1, n2 * tk_len2, negative_prompt_embeds.shape[-1])
+                    )
+                    if n2 > 1:
+                        states_list = [negative_prompt_embeds[:, 0].unsqueeze(1)]
+                        for i in range(
+                            1,
+                            max_token_length,
+                            tokenizers[text_encoder_idx].model_max_length,
+                        ):
+                            states_list.append(
+                                negative_prompt_embeds[
+                                    :,
+                                    i : i
+                                    + tokenizers[text_encoder_idx].model_max_length
+                                    - 2,
+                                ]
+                            )
+                        states_list.append(negative_prompt_embeds[:, -1].unsqueeze(1))
+                        negative_prompt_embeds = torch.cat(states_list, dim=1)
+                        negative_pooled_prompt_embeds = negative_pooled_ouput[::n2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+                text_encoder_idx += 1
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if self.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(
+                dtype=self.text_encoder_2.dtype, device=device
+            )
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(
+            bs_embed * num_images_per_prompt, seq_len, -1
+        )
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if self.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(
+                    dtype=self.text_encoder_2.dtype, device=device
+                )
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(
+                    dtype=self.unet.dtype, device=device
+                )
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(
+                1, num_images_per_prompt, 1
+            )
+            negative_prompt_embeds = negative_prompt_embeds.view(
+                batch_size * num_images_per_prompt, seq_len, -1
+            )
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(
+            1, num_images_per_prompt
+        ).view(bs_embed * num_images_per_prompt, -1)
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(
+                1, num_images_per_prompt
+            ).view(bs_embed * num_images_per_prompt, -1)
+
+        if self.text_encoder is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+        # breakpoint()
+        return (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(
+        self, image, device, num_images_per_prompt, output_hidden_states=None
+    ):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(
+                image, output_hidden_states=True
+            ).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = (
+                uncond_image_enc_hidden_states.repeat_interleave(
+                    num_images_per_prompt, dim=0
+                )
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self,
+        ip_adapter_image,
+        ip_adapter_image_embeds,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+    ):
+        image_embeds = []
+        if do_classifier_free_guidance:
+            negative_image_embeds = []
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(
+                self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
+
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+
+                image_embeds.append(single_image_embeds[None, :])
+                if do_classifier_free_guidance:
+                    negative_image_embeds.append(single_negative_image_embeds[None, :])
+        else:
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    (
+                        single_negative_image_embeds,
+                        single_image_embeds,
+                    ) = single_image_embeds.chunk(2)
+                    negative_image_embeds.append(single_negative_image_embeds)
+                image_embeds.append(single_image_embeds)
+
+        ip_adapter_image_embeds = []
+        for i, single_image_embeds in enumerate(image_embeds):
+            single_image_embeds = torch.cat(
+                [single_image_embeds] * num_images_per_prompt, dim=0
+            )
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.cat(
+                    [negative_image_embeds[i]] * num_images_per_prompt, dim=0
+                )
+                single_image_embeds = torch.cat(
+                    [single_negative_image_embeds, single_image_embeds], dim=0
+                )
+
+            single_image_embeds = single_image_embeds.to(device=device)
+            ip_adapter_image_embeds.append(single_image_embeds)
+
+        return ip_adapter_image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(
+                f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
+            )
+
+        if callback_steps is not None and (
+            not isinstance(callback_steps, int) or callback_steps <= 0
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs
+            for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (
+            not isinstance(prompt, str) and not isinstance(prompt, list)
+        ):
+            raise ValueError(
+                f"`prompt` has to be of type `str` or `list` but is {type(prompt)}"
+            )
+        elif prompt_2 is not None and (
+            not isinstance(prompt_2, str) and not isinstance(prompt_2, list)
+        ):
+            raise ValueError(
+                f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}"
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _get_add_time_ids(
+        self,
+        original_size,
+        crops_coords_top_left,
+        target_size,
+        dtype,
+        text_encoder_projection_dim=None,
+    ):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids)
+            + text_encoder_projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    @property
+    def device(self) -> torch.device:
+        r"""
+        Returns:
+            `torch.device`: The torch device on which the pipeline is located.
+        """
+        module_names, _ = self._get_signature_keys(self)
+        modules = [getattr(self, n, None) for n in module_names]
+        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+        for module in modules:
+            return module.device
+
+        return torch.device("cpu")
+
+    @property
+    def _execution_device(self):
+        """
+        Returns the device on which the pipeline's models will be executed. After calling
+        [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from
+        Accelerate's module hooks.
+        """
+        for name, model in self.components.items():
+            if (
+                not isinstance(model, torch.nn.Module)
+                or name in self._exclude_from_cpu_offload
+            ):
+                continue
+
+            if not hasattr(model, "_hf_hook"):
+                return self.device
+            for module in model.modules():
+                if (
+                    hasattr(module, "_hf_hook")
+                    and hasattr(module._hf_hook, "execution_device")
+                    and module._hf_hook.execution_device is not None
+                ):
+                    return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                FusedAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(
+        self,
+        w: torch.Tensor,
+        embedding_dim: int = 512,
+        dtype: torch.dtype = torch.float32,
+    ) -> torch.Tensor:
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            w (`torch.Tensor`):
+                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
+            embedding_dim (`int`, *optional*, defaults to 512):
+                Dimension of the embeddings to generate.
+            dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
+                Data type of the generated embeddings.
+
+        Returns:
+            `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def denoising_end(self):
+        return self._denoising_end
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[
+            Union[
+                Callable[[int, int, Dict], None],
+                PipelineCallback,
+                MultiPipelineCallbacks,
+            ]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None)
+            if self.cross_attention_kwargs is not None
+            else None
+        )
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas
+        )
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        # breakpoint()
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat(
+                [negative_pooled_prompt_embeds, add_text_embeds], dim=0
+            )
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(
+            batch_size * num_images_per_prompt, 1
+        )
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(
+            len(timesteps) - num_inference_steps * self.scheduler.order, 0
+        )
+
+        # 8.1 Apply denoising_end
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(
+                list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))
+            )
+            timesteps = timesteps[:num_inference_steps]
+
+        # 9. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(
+                batch_size * num_images_per_prompt
+            )
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = (
+                    torch.cat([latents] * 2)
+                    if self.do_classifier_free_guidance
+                    else latents
+                )
+
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                # predict the noise residual
+                added_cond_kwargs = {
+                    "text_embeds": add_text_embeds,
+                    "time_ids": add_time_ids,
+                }
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(
+                        noise_pred,
+                        noise_pred_text,
+                        guidance_rescale=self.guidance_rescale,
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop(
+                        "negative_prompt_embeds", negative_prompt_embeds
+                    )
+                    add_text_embeds = callback_outputs.pop(
+                        "add_text_embeds", add_text_embeds
+                    )
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop(
+                        "negative_add_time_ids", negative_add_time_ids
+                    )
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (
+                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                ):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = (
+                self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+            )
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(
+                    next(iter(self.vae.post_quant_conv.parameters())).dtype
+                )
+            elif latents.dtype != self.vae.dtype:
+                if torch.backends.mps.is_available():
+                    # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                    self.vae = self.vae.to(latents.dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = (
+                hasattr(self.vae.config, "latents_mean")
+                and self.vae.config.latents_mean is not None
+            )
+            has_latents_std = (
+                hasattr(self.vae.config, "latents_std")
+                and self.vae.config.latents_std is not None
+            )
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean)
+                    .view(1, 4, 1, 1)
+                    .to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std)
+                    .view(1, 4, 1, 1)
+                    .to(latents.device, latents.dtype)
+                )
+                latents = (
+                    latents * latents_std / self.vae.config.scaling_factor
+                    + latents_mean
+                )
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents)
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            # if self.watermark is not None:
+            #     image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return AniMemoryPipelineOutput(images=image)

scheduler/scheduler_config.json

@@ -1,5 +1,5 @@
 {
-  "_class_name": "EulerDiscreteScheduler",
+  "_class_name": "EulerAncestralDiscreteXPredScheduler",
   "_diffusers_version": "0.26.0",
   "beta_end": 0.012,
   "beta_schedule": "scaled_linear",

scheduler/scheduling_euler_ancestral_discrete_x_pred.py

@@ -0,0 +1,246 @@
+# Copyright 2024 Katherine Crowson, AniMemory Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.utils import logging
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.schedulers.scheduling_euler_ancestral_discrete import (
+    EulerAncestralDiscreteScheduler,
+    EulerAncestralDiscreteSchedulerOutput,
+    rescale_zero_terminal_snr,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class EulerAncestralDiscreteXPredScheduler(EulerAncestralDiscreteScheduler):
+    """
+    Ancestral sampling with Euler method steps. This model inherits from [`EulerAncestralDiscreteScheduler`]. Check the
+    superclass documentation for the args and returns.
+
+    For more details, see the original paper: https://arxiv.org/abs/2403.08381
+    """
+
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.0001,
+        beta_end: float = 0.02,
+        beta_schedule: str = "linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        prediction_type: str = "epsilon",
+        timestep_spacing: str = "linspace",
+        steps_offset: int = 0,
+    ):
+        super(EulerAncestralDiscreteXPredScheduler, self).__init__(
+            num_train_timesteps,
+            beta_start,
+            beta_end,
+            beta_schedule,
+            trained_betas,
+            prediction_type,
+            timestep_spacing,
+            steps_offset,
+        )
+
+        sigmas = np.array((1 - self.alphas_cumprod) ** 0.5, dtype=np.float32)
+        self.sigmas = torch.from_numpy(sigmas)
+
+    def rescale_betas_zero_snr(self):
+        self.betas = rescale_zero_terminal_snr(self.betas)
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+        sigmas = np.array((1 - self.alphas_cumprod) ** 0.5)
+        self.sigmas = torch.from_numpy(sigmas)
+
+    @property
+    def init_noise_sigma(self):
+        return 1.0
+
+    def scale_model_input(
+        self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+    ) -> torch.FloatTensor:
+        self.is_scale_input_called = True
+        # standard deviation of the initial noise distribution
+        return sample
+
+    def set_timesteps(
+        self, num_inference_steps: int, device: Union[str, torch.device] = None
+    ):
+        """
+        Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+        Args:
+            num_inference_steps (`int`):
+                the number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, optional):
+                the device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        """
+        self.num_inference_steps = num_inference_steps
+
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = np.linspace(
+                0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=float
+            )[::-1].copy()
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (
+                (np.arange(0, num_inference_steps) * step_ratio)
+                .round()[::-1]
+                .copy()
+                .astype(float)
+            )
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (
+                (np.arange(self.config.num_train_timesteps, 0, -step_ratio))
+                .round()
+                .copy()
+                .astype(float)
+            )
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+            )
+
+        sigmas = np.array((1 - self.alphas_cumprod) ** 0.5)
+        sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+        sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+
+        self.sigmas = torch.from_numpy(sigmas).to(device=device)
+        if str(device).startswith("mps"):
+            # mps does not support float64
+            self.timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32)
+        else:
+            self.timesteps = torch.from_numpy(timesteps).to(device=device)
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[EulerAncestralDiscreteSchedulerOutput, Tuple]:
+        """
+        Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+            timestep (`float`): current timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                current instance of sample being created by diffusion process.
+            generator (`torch.Generator`, optional): Random number generator.
+            return_dict (`bool`): option for returning tuple rather than EulerAncestralDiscreteSchedulerOutput class
+
+        Returns:
+            [`~schedulers.scheduling_utils.EulerAncestralDiscreteSchedulerOutput`] or `tuple`:
+            [`~schedulers.scheduling_utils.EulerAncestralDiscreteSchedulerOutput`] if `return_dict` is True, otherwise
+            a `tuple`. When returning a tuple, the first element is the sample tensor.
+
+        """
+
+        if (
+            isinstance(timestep, int)
+            or isinstance(timestep, torch.IntTensor)
+            or isinstance(timestep, torch.LongTensor)
+        ):
+            raise ValueError(
+                (
+                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                    " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+                    " one of the `scheduler.timesteps` as a timestep."
+                ),
+            )
+
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.to(self.timesteps.device)
+
+        step_index = (self.timesteps == timestep).nonzero().item()
+
+        if self.config.prediction_type == "sample":
+            pred_original_sample = model_output
+        else:
+            raise ValueError(
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+            )
+
+        sigma_t = self.sigmas[step_index]
+        sigma_s = self.sigmas[step_index + 1]
+        alpha_t = (1 - sigma_t**2) ** 0.5
+        alpha_s = (1 - sigma_s**2) ** 0.5
+
+        coef_sample = (sigma_s / sigma_t) ** 2 * alpha_t / alpha_s
+        coef_noise = (sigma_s / sigma_t) * (1 - (alpha_t / alpha_s) ** 2) ** 0.5
+        coef_x = alpha_s * (1 - alpha_t**2 / alpha_s**2) / sigma_t**2
+
+        device = model_output.device
+        noise = randn_tensor(
+            model_output.shape,
+            dtype=model_output.dtype,
+            device=device,
+            generator=generator,
+        )
+        prev_sample = (
+            coef_sample * sample + coef_x * pred_original_sample + coef_noise * noise
+        )
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return EulerAncestralDiscreteSchedulerOutput(
+            prev_sample=prev_sample, pred_original_sample=pred_original_sample
+        )
+
+    def add_noise(
+        self,
+        original_samples: torch.FloatTensor,
+        noise: torch.FloatTensor,
+        timesteps: torch.FloatTensor,
+    ) -> torch.FloatTensor:
+        # Make sure sigmas and timesteps have the same device and dtype as original_samples
+        sigmas = self.sigmas.to(
+            device=original_samples.device, dtype=original_samples.dtype
+        )
+        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+            # mps does not support float64
+            schedule_timesteps = self.timesteps.to(
+                original_samples.device, dtype=torch.float32
+            )
+            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+        else:
+            schedule_timesteps = self.timesteps.to(original_samples.device)
+            timesteps = timesteps.to(original_samples.device)
+
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < len(original_samples.shape):
+            sigma = sigma.unsqueeze(-1)
+
+        noisy_samples = original_samples + noise * sigma
+        return noisy_samples

text_encoder/animemory_t5.py

@@ -0,0 +1,81 @@
+# Copyright 2024 AniMemory Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import torch
+from safetensors.torch import load_file
+from transformers.models.t5.configuration_t5 import T5Config
+from transformers.models.t5.modeling_t5 import T5Stack
+
+
+class AniMemoryT5(torch.nn.Module):
+    def __init__(self, config: T5Config, embed_tokens=None):
+        super().__init__()
+        self.encoder = T5Stack(config, embed_tokens)
+        self.embed_tokens_encoder = torch.nn.Embedding(250002, 4096, padding_idx=1)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path,
+        subfolder="",
+        embed_tokens=None,
+        emb_name="weights.safetensors",
+        torch_dtype=torch.float16,
+    ):
+        cls.dtype = torch_dtype
+        config = T5Stack.config_class.from_pretrained(
+            pretrained_model_name_or_path, subfolder=subfolder
+        )
+        model = cls(config=config, embed_tokens=embed_tokens)
+        model.encoder = T5Stack.from_pretrained(
+            pretrained_model_name_or_path, subfolder=subfolder
+        )
+        embed_tokens_encoder_path = load_file(
+            os.path.join(pretrained_model_name_or_path, subfolder, emb_name)
+        )
+        model.embed_tokens_encoder.load_state_dict(embed_tokens_encoder_path)
+        model.encoder.to(torch_dtype)
+        model.embed_tokens_encoder.to(torch_dtype)
+        return model
+
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(
+            *args, **kwargs
+        )
+        super(AniMemoryT5, self).to(*args, **kwargs)
+        self.dtype = dtype if dtype is not None else self.dtype
+        self.device = device if device is not None else self.device
+        return self
+
+    def make_attn_mask(self, attn_mask):
+        seq_len = attn_mask.shape[1]
+        query = attn_mask.unsqueeze(1).float()
+        attn_mask = (
+            query.repeat([1, seq_len, 1]).unsqueeze(1).repeat([1, self.num_head, 1, 1])
+        )
+        attn_mask = attn_mask.view([-1, seq_len, seq_len])
+        return attn_mask
+
+    def forward(self, text, attention_mask):
+        embeddings = self.embed_tokens_encoder(text)
+        encoder_outputs = self.encoder(
+            inputs_embeds=embeddings,
+            attention_mask=attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_outputs.hidden_states[-2]
+        hidden_states = self.encoder.final_layer_norm(hidden_states)
+        return hidden_states, hidden_states

text_encoder_2/animemory_altclip.py

@@ -0,0 +1,119 @@
+# Copyright 2024 AniMemory Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import torch
+from safetensors.torch import load_file
+from transformers import CLIPTextConfig, CLIPTextModelWithProjection
+
+
+class AniMemoryAltCLip(torch.nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        self.model_hf = CLIPTextModelWithProjection(config)
+        self.linear_proj = torch.nn.Linear(in_features=1280, out_features=1280)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path,
+        subfolder="",
+        linear_proj_name="weights.safetensors",
+        torch_dtype=torch.float16,
+    ):
+        cls.dtype = torch_dtype
+        config = CLIPTextModelWithProjection.config_class.from_pretrained(
+            pretrained_model_name_or_path, subfolder=subfolder
+        )
+        model = cls(config=config)
+        model.model_hf = CLIPTextModelWithProjection.from_pretrained(
+            pretrained_model_name_or_path, subfolder=subfolder
+        )
+        linear_proj_state = load_file(
+            os.path.join(pretrained_model_name_or_path, subfolder, linear_proj_name)
+        )
+        model.linear_proj.load_state_dict(linear_proj_state)
+        return model
+
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(
+            *args, **kwargs
+        )
+        super(AniMemoryAltCLip, self).to(*args, **kwargs)
+        self.dtype = dtype if dtype is not None else self.dtype
+        self.device = device if device is not None else self.device
+        return self
+
+    def expand_mask(self, mask=None, dtype="", tgt_len=None):
+        """
+        Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+        """
+        bsz, src_len = mask.size()
+        tgt_len = tgt_len if tgt_len is not None else src_len
+
+        expanded_mask = (
+            mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+        )
+
+        inverted_mask = 1.0 - expanded_mask
+
+        return inverted_mask.masked_fill(
+            inverted_mask.to(torch.bool), torch.finfo(dtype).min
+        )
+
+    def make_attn_mask(self, attn_mask):
+        seq_len = attn_mask.shape[1]
+        query = attn_mask.unsqueeze(1).float()
+        attn_mask = (
+            query.repeat([1, seq_len, 1]).unsqueeze(1).repeat([1, self.num_head, 1, 1])
+        )
+        attn_mask = attn_mask.view([-1, seq_len, seq_len])
+        return attn_mask
+
+    def gradient_checkpointing_enable(
+        self,
+    ):
+        self.model_hf.gradient_checkpointing_enable()
+
+    def forward(self, text, attention_mask):
+        hidden_states = self.model_hf.text_model.embeddings(
+            input_ids=text, position_ids=None
+        )
+        if attention_mask is None:
+            print("Warning: attention_mask is None in altclip!")
+        new_attn_mask = (
+            self.expand_mask(attention_mask, hidden_states.dtype)
+            if attention_mask is not None
+            else None
+        )
+        encoder_outputs = self.model_hf.text_model.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=new_attn_mask,
+            causal_attention_mask=None,
+            output_attentions=False,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.model_hf.text_model.final_layer_norm(last_hidden_state)
+        last_hidden_state = (
+            last_hidden_state[torch.arange(last_hidden_state.shape[0]), 0]
+            @ self.model_hf.text_projection.weight
+        )
+        pooled_output = self.linear_proj(last_hidden_state)
+
+        extra_features = encoder_outputs.hidden_states[-2]
+        extra_features = self.model_hf.text_model.final_layer_norm(extra_features)
+        return extra_features, pooled_output

vae/modeling_movq.py

@@ -0,0 +1,539 @@
+# Copyright 2024 Kandinsky 3.0 Model Team, AniMemory Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from types import SimpleNamespace
+
+import torch
+import torch.nn as nn
+from packaging import version
+from safetensors.torch import load_file
+
+from diffusers.utils.accelerate_utils import apply_forward_hook
+
+
+def nonlinearity(x):
+    return x * torch.sigmoid(x)
+
+
+class SpatialNorm(nn.Module):
+    def __init__(
+        self,
+        f_channels,
+        zq_channels=None,
+        norm_layer=nn.GroupNorm,
+        freeze_norm_layer=False,
+        add_conv=False,
+        **norm_layer_params,
+    ):
+        super().__init__()
+        self.norm_layer = norm_layer(num_channels=f_channels, **norm_layer_params)
+        if zq_channels is not None:
+            if freeze_norm_layer:
+                for p in self.norm_layer.parameters:
+                    p.requires_grad = False
+            self.add_conv = add_conv
+            if self.add_conv:
+                self.conv = nn.Conv2d(
+                    zq_channels, zq_channels, kernel_size=3, stride=1, padding=1
+                )
+            self.conv_y = nn.Conv2d(
+                zq_channels, f_channels, kernel_size=1, stride=1, padding=0
+            )
+            self.conv_b = nn.Conv2d(
+                zq_channels, f_channels, kernel_size=1, stride=1, padding=0
+            )
+
+    def forward(self, f, zq=None):
+        norm_f = self.norm_layer(f)
+        if zq is not None:
+            f_size = f.shape[-2:]
+            if (
+                version.parse(torch.__version__) < version.parse("2.1")
+                and zq.dtype == torch.bfloat16
+            ):
+                zq = zq.to(dtype=torch.float32)
+                zq = torch.nn.functional.interpolate(zq, size=f_size, mode="nearest")
+                zq = zq.to(dtype=torch.bfloat16)
+            else:
+                zq = torch.nn.functional.interpolate(zq, size=f_size, mode="nearest")
+            if self.add_conv:
+                zq = self.conv(zq)
+            norm_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+        return norm_f
+
+
+def Normalize(in_channels, zq_ch=None, add_conv=None):
+    return SpatialNorm(
+        in_channels,
+        zq_ch,
+        norm_layer=nn.GroupNorm,
+        freeze_norm_layer=False,
+        add_conv=add_conv,
+        num_groups=32,
+        eps=1e-6,
+        affine=True,
+    )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=1, padding=1
+            )
+
+    def forward(self, x):
+        if (
+            version.parse(torch.__version__) < version.parse("2.1")
+            and x.dtype == torch.bfloat16
+        ):
+            x = x.to(dtype=torch.float32)
+            x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+            x = x.to(dtype=torch.bfloat16)
+        else:
+            x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=2, padding=0
+            )
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+        zq_ch=None,
+        add_conv=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels, zq_ch, add_conv=add_conv)
+        self.conv1 = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels, zq_ch, add_conv=add_conv)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
+                )
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(self, x, temb, zq=None):
+        h = x
+        h = self.norm1(h, zq)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h, zq)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, zq_ch=None, add_conv=False):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels, zq_ch, add_conv=add_conv)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x, zq=None):
+        h_ = x
+        h_ = self.norm(h_, zq)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w)
+        q = q.permute(0, 2, 1)
+        k = k.reshape(b, c, h * w)
+        w_ = torch.bmm(q, k)
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w)
+        w_ = w_.permute(0, 2, 1)
+        h_ = torch.bmm(v, w_)
+        h_ = h_.reshape(b, c, h, w)
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+
+    def forward(self, x):
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        zq_ch=None,
+        add_conv=False,
+        **ignorekwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            zq_ch=zq_ch,
+            add_conv=add_conv,
+        )
+        self.mid.attn_1 = AttnBlock(block_in, zq_ch, add_conv=add_conv)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            zq_ch=zq_ch,
+            add_conv=add_conv,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        zq_ch=zq_ch,
+                        add_conv=add_conv,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in, zq_ch, add_conv=add_conv))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)
+
+        # end
+        self.norm_out = Normalize(block_in, zq_ch, add_conv=add_conv)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, z, zq):
+        self.last_z_shape = z.shape
+        temb = None
+
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb, zq)
+        h = self.mid.attn_1(h, zq)
+        h = self.mid.block_2(h, temb, zq)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb, zq)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, zq)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h, zq)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+# Modified from MoVQ in https://github.com/ai-forever/Kandinsky-3/blob/main/kandinsky3/movq.py
+class MoVQ(nn.Module):
+    def __init__(self, generator_params: dict):
+        super().__init__()
+        z_channels = generator_params["z_channels"]
+        self.config = SimpleNamespace(**generator_params)
+        self.encoder = Encoder(**generator_params)
+        self.quant_conv = torch.nn.Conv2d(z_channels, z_channels, 1)
+        self.post_quant_conv = torch.nn.Conv2d(z_channels, z_channels, 1)
+        self.decoder = Decoder(zq_ch=z_channels, **generator_params)
+        self.dtype = None
+        self.device = None
+
+    @staticmethod
+    def get_model_config(pretrained_model_name_or_path, subfolder):
+        config_path = os.path.join(
+            pretrained_model_name_or_path, subfolder, "config.json"
+        )
+        assert os.path.exists(config_path), "config file not exists."
+        with open(config_path, "r") as f:
+            config = json.loads(f.read())
+        return config
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path,
+        subfolder="",
+        torch_dtype=torch.float32,
+    ):
+        config = cls.get_model_config(pretrained_model_name_or_path, subfolder)
+        model = cls(generator_params=config)
+        ckpt_path = os.path.join(
+            pretrained_model_name_or_path, subfolder, "movq_model.safetensors"
+        )
+        assert os.path.exists(
+            ckpt_path
+        ), f"ckpt path not exists, please check {ckpt_path}"
+        assert torch_dtype != torch.float16, "torch_dtype doesn't support fp16"
+        ckpt_weight = load_file(ckpt_path)
+        model.load_state_dict(ckpt_weight, strict=True)
+        model.to(dtype=torch_dtype)
+        return model
+
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(
+            *args, **kwargs
+        )
+        super(MoVQ, self).to(*args, **kwargs)
+        self.dtype = dtype if dtype is not None else self.dtype
+        self.device = device if device is not None else self.device
+        return self
+
+    @torch.no_grad()
+    @apply_forward_hook
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    @torch.no_grad()
+    @apply_forward_hook
+    def decode(self, quant):
+        decoder_input = self.post_quant_conv(quant)
+        decoded = self.decoder(decoder_input, quant)
+        return decoded