Stable unCLIP
Stable unCLIP checkpoints are finetuned from Stable Diffusion 2.1 checkpoints to condition on CLIP image embeddings. Stable unCLIP still conditions on text embeddings. Given the two separate conditionings, stable unCLIP can be used for text guided image variation. When combined with an unCLIP prior, it can also be used for full text to image generation.
The abstract from the paper is:
Contrastive models like CLIP have been shown to learn robust representations of images that capture both semantics and style. To leverage these representations for image generation, we propose a two-stage model: a prior that generates a CLIP image embedding given a text caption, and a decoder that generates an image conditioned on the image embedding. We show that explicitly generating image representations improves image diversity with minimal loss in photorealism and caption similarity. Our decoders conditioned on image representations can also produce variations of an image that preserve both its semantics and style, while varying the non-essential details absent from the image representation. Moreover, the joint embedding space of CLIP enables language-guided image manipulations in a zero-shot fashion. We use diffusion models for the decoder and experiment with both autoregressive and diffusion models for the prior, finding that the latter are computationally more efficient and produce higher-quality samples.
Tips
Stable unCLIP takes noise_level as input during inference which determines how much noise is added to the image embeddings. A higher noise_level increases variation in the final un-noised images. By default, we do not add any additional noise to the image embeddings (noise_level = 0).
Text-to-Image Generation
Stable unCLIP can be leveraged for text-to-image generation by pipelining it with the prior model of KakaoBrain’s open source DALL-E 2 replication Karlo:
import torch
from diffusers import UnCLIPScheduler, DDPMScheduler, StableUnCLIPPipeline
from diffusers.models import PriorTransformer
from transformers import CLIPTokenizer, CLIPTextModelWithProjection
prior_model_id = "kakaobrain/karlo-v1-alpha"
data_type = torch.float16
prior = PriorTransformer.from_pretrained(prior_model_id, subfolder="prior", torch_dtype=data_type)
prior_text_model_id = "openai/clip-vit-large-patch14"
prior_tokenizer = CLIPTokenizer.from_pretrained(prior_text_model_id)
prior_text_model = CLIPTextModelWithProjection.from_pretrained(prior_text_model_id, torch_dtype=data_type)
prior_scheduler = UnCLIPScheduler.from_pretrained(prior_model_id, subfolder="prior_scheduler")
prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config)
stable_unclip_model_id = "stabilityai/stable-diffusion-2-1-unclip-small"
pipe = StableUnCLIPPipeline.from_pretrained(
stable_unclip_model_id,
torch_dtype=data_type,
variant="fp16",
prior_tokenizer=prior_tokenizer,
prior_text_encoder=prior_text_model,
prior=prior,
prior_scheduler=prior_scheduler,
)
pipe = pipe.to("cuda")
wave_prompt = "dramatic wave, the Oceans roar, Strong wave spiral across the oceans as the waves unfurl into roaring crests; perfect wave form; perfect wave shape; dramatic wave shape; wave shape unbelievable; wave; wave shape spectacular"
image = pipe(prompt=wave_prompt).images[0]
imageFor text-to-image we use stabilityai/stable-diffusion-2-1-unclip-small as it was trained on CLIP ViT-L/14 embedding, the same as the Karlo model prior. stabilityai/stable-diffusion-2-1-unclip was trained on OpenCLIP ViT-H, so we don’t recommend its use.
Text guided Image-to-Image Variation
from diffusers import StableUnCLIPImg2ImgPipeline
from diffusers.utils import load_image
import torch
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1-unclip", torch_dtype=torch.float16, variation="fp16"
)
pipe = pipe.to("cuda")
url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/tarsila_do_amaral.png"
init_image = load_image(url)
images = pipe(init_image).images
images[0].save("variation_image.png")Optionally, you can also pass a prompt to pipe such as:
prompt = "A fantasy landscape, trending on artstation"
image = pipe(init_image, prompt=prompt).images[0]
imageMake sure to check out the Schedulers guide to learn how to explore the tradeoff between scheduler speed and quality, and see the reuse components across pipelines section to learn how to efficiently load the same components into multiple pipelines.
StableUnCLIPPipeline
class diffusers.StableUnCLIPPipeline
< source >( prior_tokenizer: CLIPTokenizer prior_text_encoder: CLIPTextModelWithProjection prior: PriorTransformer prior_scheduler: KarrasDiffusionSchedulers image_normalizer: StableUnCLIPImageNormalizer image_noising_scheduler: KarrasDiffusionSchedulers tokenizer: CLIPTokenizer text_encoder: CLIPTextModelWithProjection unet: UNet2DConditionModel scheduler: KarrasDiffusionSchedulers vae: AutoencoderKL )
Parameters
- prior_tokenizer (
CLIPTokenizer) — ACLIPTokenizer. - prior_text_encoder (
CLIPTextModelWithProjection) — FrozenCLIPTextModelWithProjectiontext-encoder. - prior (PriorTransformer) — The canonical unCLIP prior to approximate the image embedding from the text embedding.
- prior_scheduler (
KarrasDiffusionSchedulers) — Scheduler used in the prior denoising process. - image_normalizer (
StableUnCLIPImageNormalizer) — Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image embeddings after the noise has been applied. - image_noising_scheduler (
KarrasDiffusionSchedulers) — Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined by thenoise_level. - tokenizer (
CLIPTokenizer) — ACLIPTokenizer. - text_encoder (
CLIPTextModel) — FrozenCLIPTextModeltext-encoder. - unet (UNet2DConditionModel) — A UNet2DConditionModel to denoise the encoded image latents.
- scheduler (
KarrasDiffusionSchedulers) — A scheduler to be used in combination withunetto denoise the encoded image latents. - vae (AutoencoderKL) — Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
Pipeline for text-to-image generation using stable unCLIP.
This model inherits from DiffusionPipeline. Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.).
The pipeline also inherits the following loading methods:
- load_textual_inversion() for loading textual inversion embeddings
- load_lora_weights() for loading LoRA weights
- save_lora_weights() for saving LoRA weights
__call__
< source >( prompt: typing.Union[str, typing.List[str], NoneType] = None height: typing.Optional[int] = None width: typing.Optional[int] = None num_inference_steps: int = 20 guidance_scale: float = 10.0 negative_prompt: typing.Union[str, typing.List[str], NoneType] = None num_images_per_prompt: typing.Optional[int] = 1 eta: float = 0.0 generator: typing.Optional[torch._C.Generator] = None latents: typing.Optional[torch.Tensor] = None prompt_embeds: typing.Optional[torch.Tensor] = None negative_prompt_embeds: typing.Optional[torch.Tensor] = None output_type: typing.Optional[str] = 'pil' return_dict: bool = True callback: typing.Optional[typing.Callable[[int, int, torch.Tensor], NoneType]] = None callback_steps: int = 1 cross_attention_kwargs: typing.Optional[typing.Dict[str, typing.Any]] = None noise_level: int = 0 prior_num_inference_steps: int = 25 prior_guidance_scale: float = 4.0 prior_latents: typing.Optional[torch.Tensor] = None clip_skip: typing.Optional[int] = None ) → ImagePipelineOutput or tuple
Parameters
- prompt (
strorList[str], optional) — The prompt or prompts to guide image generation. If not defined, you need to passprompt_embeds. - height (
int, optional, defaults toself.unet.config.sample_size * self.vae_scale_factor) — The height in pixels of the generated image. - width (
int, optional, defaults toself.unet.config.sample_size * self.vae_scale_factor) — The width in pixels of the generated image. - num_inference_steps (
int, optional, defaults to 20) — The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. - guidance_scale (
float, optional, defaults to 10.0) — A higher guidance scale value encourages the model to generate images closely linked to the textpromptat the expense of lower image quality. Guidance scale is enabled whenguidance_scale > 1. - negative_prompt (
strorList[str], optional) — The prompt or prompts to guide what to not include in image generation. If not defined, you need to passnegative_prompt_embedsinstead. Ignored when not using guidance (guidance_scale < 1). - 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) from the DDIM paper. Only applies to the DDIMScheduler, and is ignored in other schedulers. - generator (
torch.GeneratororList[torch.Generator], optional) — Atorch.Generatorto 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 is generated by sampling using the supplied randomgenerator. - prompt_embeds (
torch.Tensor, optional) — Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, text embeddings are generated from thepromptinput argument. - negative_prompt_embeds (
torch.Tensor, optional) — Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided,negative_prompt_embedsare generated from thenegative_promptinput argument. - output_type (
str, optional, defaults to"pil") — The output format of the generated image. Choose betweenPIL.Imageornp.array. - return_dict (
bool, optional, defaults toTrue) — Whether or not to return a ImagePipelineOutput instead of a plain tuple. - callback (
Callable, optional) — A function that calls everycallback_stepssteps during inference. The function is called with the following arguments:callback(step: int, timestep: int, latents: torch.Tensor). - callback_steps (
int, optional, defaults to 1) — The frequency at which thecallbackfunction is called. If not specified, the callback is called at every step. - cross_attention_kwargs (
dict, optional) — A kwargs dictionary that if specified is passed along to theAttentionProcessoras defined inself.processor. - noise_level (
int, optional, defaults to0) — The amount of noise to add to the image embeddings. A highernoise_levelincreases the variance in the final un-noised images. See StableUnCLIPPipeline.noise_image_embeddings() for more details. - prior_num_inference_steps (
int, optional, defaults to 25) — The number of denoising steps in the prior denoising process. More denoising steps usually lead to a higher quality image at the expense of slower inference. - prior_guidance_scale (
float, optional, defaults to 4.0) — A higher guidance scale value encourages the model to generate images closely linked to the textpromptat the expense of lower image quality. Guidance scale is enabled whenguidance_scale > 1. - prior_latents (
torch.Tensor, optional) — Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image embedding generation in the prior denoising process. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor is generated by sampling using the supplied randomgenerator. - 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.
Returns
ImagePipelineOutput or tuple
~ pipeline_utils.ImagePipelineOutput if return_dict is True, otherwise a tuple. When returning
a tuple, the first element is a list with the generated images.
The call function to the pipeline for generation.
Examples:
>>> import torch
>>> from diffusers import StableUnCLIPPipeline
>>> pipe = StableUnCLIPPipeline.from_pretrained(
... "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16
... ) # TODO update model path
>>> pipe = pipe.to("cuda")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> images = pipe(prompt).images
>>> images[0].save("astronaut_horse.png")enable_attention_slicing
< source >( slice_size: typing.Union[int, str, NoneType] = 'auto' )
Parameters
- slice_size (
strorint, optional, defaults to"auto") — When"auto", halves the input to the attention heads, so attention will be computed in two steps. If"max", maximum amount of memory will be saved by running only one slice at a time. If a number is provided, uses as many slices asattention_head_dim // slice_size. In this case,attention_head_dimmust be a multiple ofslice_size.
Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor in slices to compute attention in several steps. For more than one attention head, the computation is performed sequentially over each head. This is useful to save some memory in exchange for a small speed decrease.
⚠️ Don’t enable attention slicing if you’re already using scaled_dot_product_attention (SDPA) from PyTorch
2.0 or xFormers. These attention computations are already very memory efficient so you won’t need to enable
this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs!
Examples:
>>> import torch
>>> from diffusers import StableDiffusionPipeline
>>> pipe = StableDiffusionPipeline.from_pretrained(
... "runwayml/stable-diffusion-v1-5",
... torch_dtype=torch.float16,
... use_safetensors=True,
... )
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> pipe.enable_attention_slicing()
>>> image = pipe(prompt).images[0]Disable sliced attention computation. If enable_attention_slicing was previously called, attention is
computed in one step.
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
Disable sliced VAE decoding. If enable_vae_slicing was previously enabled, this method will go back to
computing decoding in one step.
enable_xformers_memory_efficient_attention
< source >( attention_op: typing.Optional[typing.Callable] = None )
Parameters
- attention_op (
Callable, optional) — Override the defaultNoneoperator for use asopargument to thememory_efficient_attention()function of xFormers.
Enable memory efficient attention from xFormers. When this option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed up during training is not guaranteed.
⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes precedent.
Examples:
>>> import torch
>>> from diffusers import DiffusionPipeline
>>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
>>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
>>> pipe = pipe.to("cuda")
>>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
>>> # Workaround for not accepting attention shape using VAE for Flash Attention
>>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None)Disable memory efficient attention from xFormers.
encode_prompt
< source >( prompt device num_images_per_prompt do_classifier_free_guidance negative_prompt = None prompt_embeds: typing.Optional[torch.Tensor] = None negative_prompt_embeds: typing.Optional[torch.Tensor] = None lora_scale: typing.Optional[float] = None clip_skip: typing.Optional[int] = None )
Parameters
- prompt (
strorList[str], optional) — prompt to be encoded 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 (
strorList[str], optional) — The prompt or prompts not to guide the image generation. If not defined, one has to passnegative_prompt_embedsinstead. Ignored when not using guidance (i.e., ignored ifguidance_scaleis less than1). - 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 frompromptinput 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 fromnegative_promptinput 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.
Encodes the prompt into text encoder hidden states.
noise_image_embeddings
< source >( image_embeds: Tensor noise_level: int noise: typing.Optional[torch.Tensor] = None generator: typing.Optional[torch._C.Generator] = None )
Add noise to the image embeddings. The amount of noise is controlled by a noise_level input. A higher
noise_level increases the variance in the final un-noised images.
The noise is applied in two ways:
- A noise schedule is applied directly to the embeddings.
- A vector of sinusoidal time embeddings are appended to the output.
In both cases, the amount of noise is controlled by the same noise_level.
The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
StableUnCLIPImg2ImgPipeline
class diffusers.StableUnCLIPImg2ImgPipeline
< source >( feature_extractor: CLIPImageProcessor image_encoder: CLIPVisionModelWithProjection image_normalizer: StableUnCLIPImageNormalizer image_noising_scheduler: KarrasDiffusionSchedulers tokenizer: CLIPTokenizer text_encoder: CLIPTextModel unet: UNet2DConditionModel scheduler: KarrasDiffusionSchedulers vae: AutoencoderKL )
Parameters
- feature_extractor (
CLIPImageProcessor) — Feature extractor for image pre-processing before being encoded. - image_encoder (
CLIPVisionModelWithProjection) — CLIP vision model for encoding images. - image_normalizer (
StableUnCLIPImageNormalizer) — Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image embeddings after the noise has been applied. - image_noising_scheduler (
KarrasDiffusionSchedulers) — Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined by thenoise_level. - tokenizer (
~transformers.CLIPTokenizer) — A [~transformers.CLIPTokenizer)]. - text_encoder (CLIPTextModel) — Frozen CLIPTextModel text-encoder.
- unet (UNet2DConditionModel) — A UNet2DConditionModel to denoise the encoded image latents.
- scheduler (
KarrasDiffusionSchedulers) — A scheduler to be used in combination withunetto denoise the encoded image latents. - vae (AutoencoderKL) — Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
Pipeline for text-guided image-to-image generation using stable unCLIP.
This model inherits from DiffusionPipeline. Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.).
The pipeline also inherits the following loading methods:
- load_textual_inversion() for loading textual inversion embeddings
- load_lora_weights() for loading LoRA weights
- save_lora_weights() for saving LoRA weights
__call__
< source >( image: typing.Union[torch.Tensor, PIL.Image.Image] = None prompt: typing.Union[str, typing.List[str]] = None height: typing.Optional[int] = None width: typing.Optional[int] = None num_inference_steps: int = 20 guidance_scale: float = 10 negative_prompt: typing.Union[str, typing.List[str], NoneType] = None num_images_per_prompt: typing.Optional[int] = 1 eta: float = 0.0 generator: typing.Optional[torch._C.Generator] = None latents: typing.Optional[torch.Tensor] = None prompt_embeds: typing.Optional[torch.Tensor] = None negative_prompt_embeds: typing.Optional[torch.Tensor] = None output_type: typing.Optional[str] = 'pil' return_dict: bool = True callback: typing.Optional[typing.Callable[[int, int, torch.Tensor], NoneType]] = None callback_steps: int = 1 cross_attention_kwargs: typing.Optional[typing.Dict[str, typing.Any]] = None noise_level: int = 0 image_embeds: typing.Optional[torch.Tensor] = None clip_skip: typing.Optional[int] = None ) → ImagePipelineOutput or tuple
Parameters
- prompt (
strorList[str], optional) — The prompt or prompts to guide the image generation. If not defined, eitherprompt_embedswill be used or prompt is initialized to"". - image (
torch.TensororPIL.Image.Image) —Imageor tensor representing an image batch. The image is encoded to its CLIP embedding which theunetis conditioned on. The image is not encoded by thevaeand then used as the latents in the denoising process like it is in the standard Stable Diffusion text-guided image variation process. - height (
int, optional, defaults toself.unet.config.sample_size * self.vae_scale_factor) — The height in pixels of the generated image. - width (
int, optional, defaults toself.unet.config.sample_size * self.vae_scale_factor) — The width in pixels of the generated image. - num_inference_steps (
int, optional, defaults to 20) — The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. - guidance_scale (
float, optional, defaults to 10.0) — A higher guidance scale value encourages the model to generate images closely linked to the textpromptat the expense of lower image quality. Guidance scale is enabled whenguidance_scale > 1. - negative_prompt (
strorList[str], optional) — The prompt or prompts to guide what to not include in image generation. If not defined, you need to passnegative_prompt_embedsinstead. Ignored when not using guidance (guidance_scale < 1). - 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) from the DDIM paper. Only applies to the DDIMScheduler, and is ignored in other schedulers. - generator (
torch.GeneratororList[torch.Generator], optional) — Atorch.Generatorto 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 is generated by sampling using the supplied randomgenerator. - prompt_embeds (
torch.Tensor, optional) — Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, text embeddings are generated from thepromptinput argument. - negative_prompt_embeds (
torch.Tensor, optional) — Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided,negative_prompt_embedsare generated from thenegative_promptinput argument. - output_type (
str, optional, defaults to"pil") — The output format of the generated image. Choose betweenPIL.Imageornp.array. - return_dict (
bool, optional, defaults toTrue) — Whether or not to return a ImagePipelineOutput instead of a plain tuple. - callback (
Callable, optional) — A function that calls everycallback_stepssteps during inference. The function is called with the following arguments:callback(step: int, timestep: int, latents: torch.Tensor). - callback_steps (
int, optional, defaults to 1) — The frequency at which thecallbackfunction is called. If not specified, the callback is called at every step. - cross_attention_kwargs (
dict, optional) — A kwargs dictionary that if specified is passed along to theAttentionProcessoras defined inself.processor. - noise_level (
int, optional, defaults to0) — The amount of noise to add to the image embeddings. A highernoise_levelincreases the variance in the final un-noised images. See StableUnCLIPPipeline.noise_image_embeddings() for more details. - image_embeds (
torch.Tensor, optional) — Pre-generated CLIP embeddings to condition theuneton. These latents are not used in the denoising process. If you want to provide pre-generated latents, pass them to__call__aslatents. - 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.
Returns
ImagePipelineOutput or tuple
~ pipeline_utils.ImagePipelineOutput if return_dict is True, otherwise a tuple. When returning
a tuple, the first element is a list with the generated images.
The call function to the pipeline for generation.
Examples:
>>> import requests
>>> import torch
>>> from PIL import Image
>>> from io import BytesIO
>>> from diffusers import StableUnCLIPImg2ImgPipeline
>>> pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
... "stabilityai/stable-diffusion-2-1-unclip-small", torch_dtype=torch.float16
... )
>>> pipe = pipe.to("cuda")
>>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
>>> response = requests.get(url)
>>> init_image = Image.open(BytesIO(response.content)).convert("RGB")
>>> init_image = init_image.resize((768, 512))
>>> prompt = "A fantasy landscape, trending on artstation"
>>> images = pipe(init_image, prompt).images
>>> images[0].save("fantasy_landscape.png")enable_attention_slicing
< source >( slice_size: typing.Union[int, str, NoneType] = 'auto' )
Parameters
- slice_size (
strorint, optional, defaults to"auto") — When"auto", halves the input to the attention heads, so attention will be computed in two steps. If"max", maximum amount of memory will be saved by running only one slice at a time. If a number is provided, uses as many slices asattention_head_dim // slice_size. In this case,attention_head_dimmust be a multiple ofslice_size.
Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor in slices to compute attention in several steps. For more than one attention head, the computation is performed sequentially over each head. This is useful to save some memory in exchange for a small speed decrease.
⚠️ Don’t enable attention slicing if you’re already using scaled_dot_product_attention (SDPA) from PyTorch
2.0 or xFormers. These attention computations are already very memory efficient so you won’t need to enable
this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs!
Examples:
>>> import torch
>>> from diffusers import StableDiffusionPipeline
>>> pipe = StableDiffusionPipeline.from_pretrained(
... "runwayml/stable-diffusion-v1-5",
... torch_dtype=torch.float16,
... use_safetensors=True,
... )
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> pipe.enable_attention_slicing()
>>> image = pipe(prompt).images[0]Disable sliced attention computation. If enable_attention_slicing was previously called, attention is
computed in one step.
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
Disable sliced VAE decoding. If enable_vae_slicing was previously enabled, this method will go back to
computing decoding in one step.
enable_xformers_memory_efficient_attention
< source >( attention_op: typing.Optional[typing.Callable] = None )
Parameters
- attention_op (
Callable, optional) — Override the defaultNoneoperator for use asopargument to thememory_efficient_attention()function of xFormers.
Enable memory efficient attention from xFormers. When this option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed up during training is not guaranteed.
⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes precedent.
Examples:
>>> import torch
>>> from diffusers import DiffusionPipeline
>>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
>>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
>>> pipe = pipe.to("cuda")
>>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
>>> # Workaround for not accepting attention shape using VAE for Flash Attention
>>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None)Disable memory efficient attention from xFormers.
encode_prompt
< source >( prompt device num_images_per_prompt do_classifier_free_guidance negative_prompt = None prompt_embeds: typing.Optional[torch.Tensor] = None negative_prompt_embeds: typing.Optional[torch.Tensor] = None lora_scale: typing.Optional[float] = None clip_skip: typing.Optional[int] = None )
Parameters
- prompt (
strorList[str], optional) — prompt to be encoded 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 (
strorList[str], optional) — The prompt or prompts not to guide the image generation. If not defined, one has to passnegative_prompt_embedsinstead. Ignored when not using guidance (i.e., ignored ifguidance_scaleis less than1). - 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 frompromptinput 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 fromnegative_promptinput 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.
Encodes the prompt into text encoder hidden states.
noise_image_embeddings
< source >( image_embeds: Tensor noise_level: int noise: typing.Optional[torch.Tensor] = None generator: typing.Optional[torch._C.Generator] = None )
Add noise to the image embeddings. The amount of noise is controlled by a noise_level input. A higher
noise_level increases the variance in the final un-noised images.
The noise is applied in two ways:
- A noise schedule is applied directly to the embeddings.
- A vector of sinusoidal time embeddings are appended to the output.
In both cases, the amount of noise is controlled by the same noise_level.
The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
ImagePipelineOutput
class diffusers.ImagePipelineOutput
< source >( images: typing.Union[typing.List[PIL.Image.Image], numpy.ndarray] )
Output class for image pipelines.