Upload 17 files

.gitattributes

@@ -1,35 +1,35 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,6 @@
+*.pt
+*.yaml
+**/__pycache__
+*.pyc
+
+venv/

README.md

@@ -0,0 +1,29 @@
+---
+license: openrail
+pipeline_tag: image-to-3d
+---
+
+This is a copy of [ashawkey/imagedream-ipmv-diffusers](https://huggingface.co/ashawkey/imagedream-ipmv-diffusers).
+
+It is hosted here for persistence throughout the ML for 3D course.
+
+# MVDream-diffusers Model Card
+
+This is a port of https://huggingface.co/Peng-Wang/ImageDream into diffusers.
+
+For usage, please check: https://github.com/ashawkey/mvdream_diffusers
+
+## Citation
+
+```
+@article{wang2023imagedream,
+  title={ImageDream: Image-Prompt Multi-view Diffusion for 3D Generation},
+  author={Wang, Peng and Shi, Yichun},
+  journal={arXiv preprint arXiv:2312.02201},
+  year={2023}
+}
+```
+
+## Misuse, Malicious Use, and Out-of-Scope Use
+
+The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.

feature_extractor/preprocessor_config.json

@@ -0,0 +1,29 @@
+{
+  "crop_size": {
+    "height": 224,
+    "width": 224
+  },
+  "do_center_crop": true,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "feature_extractor_type": "CLIPFeatureExtractor",
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "CLIPImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "shortest_edge": 224
+  },
+  "use_square_size": false
+}

image_encoder/config.json

@@ -0,0 +1,23 @@
+{
+  "_name_or_path": "laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
+  "architectures": [
+    "CLIPVisionModel"
+  ],
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 32,
+  "patch_size": 14,
+  "projection_dim": 1024,
+  "torch_dtype": "float16",
+  "transformers_version": "4.35.2"
+}

model_index.json

@@ -0,0 +1,33 @@
+{
+  "_class_name": "MVDreamPipeline",
+  "_diffusers_version": "0.25.0",
+  "feature_extractor": [
+    "transformers",
+    "CLIPImageProcessor"
+  ],
+  "image_encoder": [
+    "transformers",
+    "CLIPVisionModel"
+  ],
+  "requires_safety_checker": false,
+  "scheduler": [
+    "diffusers",
+    "DDIMScheduler"
+  ],
+  "text_encoder": [
+    "transformers",
+    "CLIPTextModel"
+  ],
+  "tokenizer": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "unet": [
+    "mv_unet",
+    "MultiViewUNetModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}

pipeline.py

@@ -0,0 +1,1583 @@
+import inspect
+import math
+from inspect import isfunction
+from typing import Any, Callable, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# require xformers!
+import xformers
+import xformers.ops
+from diffusers import AutoencoderKL, DiffusionPipeline
+from diffusers.configuration_utils import ConfigMixin, FrozenDict
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.schedulers import DDIMScheduler
+from diffusers.utils import (deprecate, is_accelerate_available,
+                             is_accelerate_version, logging)
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange, repeat
+from kiui.cam import orbit_camera
+from transformers import (CLIPImageProcessor, CLIPTextModel, CLIPTokenizer,
+                          CLIPVisionModel)
+
+
+def get_camera(
+    num_frames,
+    elevation=15,
+    azimuth_start=0,
+    azimuth_span=360,
+    blender_coord=True,
+    extra_view=False,
+):
+    angle_gap = azimuth_span / num_frames
+    cameras = []
+    for azimuth in np.arange(azimuth_start, azimuth_span + azimuth_start, angle_gap):
+
+        pose = orbit_camera(
+            -elevation, azimuth, radius=1
+        )  # kiui's elevation is negated, [4, 4]
+
+        # opengl to blender
+        if blender_coord:
+            pose[2] *= -1
+            pose[[1, 2]] = pose[[2, 1]]
+
+        cameras.append(pose.flatten())
+
+    if extra_view:
+        cameras.append(np.zeros_like(cameras[0]))
+
+    return torch.from_numpy(np.stack(cameras, axis=0)).float()  # [num_frames, 16]
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period)
+            * torch.arange(start=0, end=half, dtype=torch.float32)
+            / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None] * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat(
+                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
+            )
+    else:
+        embedding = repeat(timesteps, "b -> b d", d=dim)
+    # import pdb; pdb.set_trace()
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def default(val, d):
+    if val is not None:
+        return val
+    return d() if isfunction(d) else d
+
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = (
+            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
+            if not glu
+            else GEGLU(dim, inner_dim)
+        )
+
+        self.net = nn.Sequential(
+            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class MemoryEfficientCrossAttention(nn.Module):
+    # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    def __init__(
+        self,
+        query_dim,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        ip_dim=0,
+        ip_weight=1,
+    ):
+        super().__init__()
+
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.ip_dim = ip_dim
+        self.ip_weight = ip_weight
+
+        if self.ip_dim > 0:
+            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
+        )
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x, context=None):
+        q = self.to_q(x)
+        context = default(context, x)
+
+        if self.ip_dim > 0:
+            # context： [B, 77 + 16(ip), 1024]
+            token_len = context.shape[1]
+            context_ip = context[:, -self.ip_dim :, :]
+            k_ip = self.to_k_ip(context_ip)
+            v_ip = self.to_v_ip(context_ip)
+            context = context[:, : (token_len - self.ip_dim), :]
+
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        out = xformers.ops.memory_efficient_attention(
+            q, k, v, attn_bias=None, op=self.attention_op
+        )
+
+        if self.ip_dim > 0:
+            k_ip, v_ip = map(
+                lambda t: t.unsqueeze(3)
+                .reshape(b, t.shape[1], self.heads, self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b * self.heads, t.shape[1], self.dim_head)
+                .contiguous(),
+                (k_ip, v_ip),
+            )
+            # actually compute the attention, what we cannot get enough of
+            out_ip = xformers.ops.memory_efficient_attention(
+                q, k_ip, v_ip, attn_bias=None, op=self.attention_op
+            )
+            out = out + self.ip_weight * out_ip
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        return self.to_out(out)
+
+
+class BasicTransformerBlock3D(nn.Module):
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        context_dim,
+        dropout=0.0,
+        gated_ff=True,
+        ip_dim=0,
+        ip_weight=1,
+    ):
+        super().__init__()
+
+        self.attn1 = MemoryEfficientCrossAttention(
+            query_dim=dim,
+            context_dim=None,  # self-attention
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+        )
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = MemoryEfficientCrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            # ip only applies to cross-attention
+            ip_dim=ip_dim,
+            ip_weight=ip_weight,
+        )
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+
+    def forward(self, x, context=None, num_frames=1):
+        x = rearrange(x, "(b f) l c -> b (f l) c", f=num_frames).contiguous()
+        x = self.attn1(self.norm1(x), context=None) + x
+        x = rearrange(x, "b (f l) c -> (b f) l c", f=num_frames).contiguous()
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        context_dim,  # cross attention input dim
+        depth=1,
+        dropout=0.0,
+        ip_dim=0,
+        ip_weight=1,
+    ):
+        super().__init__()
+
+        if not isinstance(context_dim, list):
+            context_dim = [context_dim]
+
+        self.in_channels = in_channels
+
+        inner_dim = n_heads * d_head
+        self.norm = nn.GroupNorm(
+            num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock3D(
+                    inner_dim,
+                    n_heads,
+                    d_head,
+                    context_dim=context_dim[d],
+                    dropout=dropout,
+                    ip_dim=ip_dim,
+                    ip_weight=ip_weight,
+                )
+                for d in range(depth)
+            ]
+        )
+
+        self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
+
+    def forward(self, x, context=None, num_frames=1):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context]
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = rearrange(x, "b c h w -> b (h w) c").contiguous()
+        x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context[i], num_frames=num_frames)
+        x = self.proj_out(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous()
+
+        return x + x_in
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, h, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q, k, v = map(
+            lambda t: t.reshape(b, t.shape[1], self.heads, -1)
+            .transpose(1, 2)
+            .reshape(b, self.heads, t.shape[1], -1)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(
+            -2, -1
+        )  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, h, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+    ):
+        super().__init__()
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+        self.proj_in = nn.Linear(embedding_dim, dim)
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        nn.Sequential(
+                            nn.LayerNorm(dim),
+                            nn.Linear(dim, dim * ff_mult, bias=False),
+                            nn.GELU(),
+                            nn.Linear(dim * ff_mult, dim, bias=False),
+                        ),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1)
+        x = self.proj_in(x)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+class CondSequential(nn.Sequential):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, num_frames=1):
+        for layer in self:
+            if isinstance(layer, ResBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer3D):
+                x = layer(x, context, num_frames=num_frames)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(
+                dims, self.channels, self.out_channels, 3, padding=padding
+            )
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims,
+                self.channels,
+                self.out_channels,
+                3,
+                stride=stride,
+                padding=padding,
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(nn.Module):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            nn.GroupNorm(32, channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            nn.GroupNorm(32, self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class MultiViewUNetModel(ModelMixin, ConfigMixin):
+    """
+    The full multi-view UNet model with attention, timestep embedding and camera embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    :param camera_dim: dimensionality of camera input.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        transformer_depth=1,
+        context_dim=None,
+        n_embed=None,
+        num_attention_blocks=None,
+        adm_in_channels=None,
+        camera_dim=None,
+        ip_dim=0,  # imagedream uses ip_dim > 0
+        ip_weight=1.0,
+        **kwargs,
+    ):
+        super().__init__()
+        assert context_dim is not None
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert (
+                num_head_channels != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        if num_head_channels == -1:
+            assert (
+                num_heads != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError(
+                    "provide num_res_blocks either as an int (globally constant) or "
+                    "as a list/tuple (per-level) with the same length as channel_mult"
+                )
+            self.num_res_blocks = num_res_blocks
+
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(
+                map(
+                    lambda i: self.num_res_blocks[i] >= num_attention_blocks[i],
+                    range(len(num_attention_blocks)),
+                )
+            )
+            print(
+                f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                f"attention will still not be set."
+            )
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        self.ip_dim = ip_dim
+        self.ip_weight = ip_weight
+
+        if self.ip_dim > 0:
+            self.image_embed = Resampler(
+                dim=context_dim,
+                depth=4,
+                dim_head=64,
+                heads=12,
+                num_queries=ip_dim,  # num token
+                embedding_dim=1280,
+                output_dim=context_dim,
+                ff_mult=4,
+            )
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            nn.Linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim),
+        )
+
+        if camera_dim is not None:
+            time_embed_dim = model_channels * 4
+            self.camera_embed = nn.Sequential(
+                nn.Linear(camera_dim, time_embed_dim),
+                nn.SiLU(),
+                nn.Linear(time_embed_dim, time_embed_dim),
+            )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(self.num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                # print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        nn.Linear(adm_in_channels, time_embed_dim),
+                        nn.SiLU(),
+                        nn.Linear(time_embed_dim, time_embed_dim),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [CondSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers: List[Any] = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+
+                    if num_attention_blocks is None or nr < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer3D(
+                                ch,
+                                num_heads,
+                                dim_head,
+                                context_dim=context_dim,
+                                depth=transformer_depth,
+                                ip_dim=self.ip_dim,
+                                ip_weight=self.ip_weight,
+                            )
+                        )
+                self.input_blocks.append(CondSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    CondSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+
+        self.middle_block = CondSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            SpatialTransformer3D(
+                ch,
+                num_heads,
+                dim_head,
+                context_dim=context_dim,
+                depth=transformer_depth,
+                ip_dim=self.ip_dim,
+                ip_weight=self.ip_weight,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+
+                    if num_attention_blocks is None or i < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer3D(
+                                ch,
+                                num_heads,
+                                dim_head,
+                                context_dim=context_dim,
+                                depth=transformer_depth,
+                                ip_dim=self.ip_dim,
+                                ip_weight=self.ip_weight,
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(CondSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            nn.GroupNorm(32, ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+                nn.GroupNorm(32, ch),
+                conv_nd(dims, model_channels, n_embed, 1),
+                # nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+            )
+
+    def forward(
+        self,
+        x,
+        timesteps=None,
+        context=None,
+        y=None,
+        camera=None,
+        num_frames=1,
+        ip=None,
+        ip_img=None,
+        **kwargs,
+    ):
+        """
+        Apply the model to an input batch.
+        :param x: an [(N x F) x C x ...] Tensor of inputs. F is the number of frames (views).
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :param num_frames: a integer indicating number of frames for tensor reshaping.
+        :return: an [(N x F) x C x ...] Tensor of outputs. F is the number of frames (views).
+        """
+        assert (
+            x.shape[0] % num_frames == 0
+        ), "input batch size must be dividable by num_frames!"
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+
+        hs = []
+
+        t_emb = timestep_embedding(
+            timesteps, self.model_channels, repeat_only=False
+        ).to(x.dtype)
+
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y is not None
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        # Add camera embeddings
+        if camera is not None:
+            emb = emb + self.camera_embed(camera)
+
+        # imagedream variant
+        if self.ip_dim > 0:
+            x[(num_frames - 1) :: num_frames, :, :, :] = ip_img  # place at [4, 9]
+            ip_emb = self.image_embed(ip)
+            context = torch.cat((context, ip_emb), 1)
+
+        h = x
+        for module in self.input_blocks:
+            h = module(h, emb, context, num_frames=num_frames)
+            hs.append(h)
+        h = self.middle_block(h, emb, context, num_frames=num_frames)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context, num_frames=num_frames)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class MVDreamPipeline(DiffusionPipeline):
+
+    _optional_components = ["feature_extractor", "image_encoder"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        unet: MultiViewUNetModel,
+        tokenizer: CLIPTokenizer,
+        text_encoder: CLIPTextModel,
+        scheduler: DDIMScheduler,
+        # imagedream variant
+        feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModel,
+        requires_safety_checker: bool = False,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:  # type: ignore
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "  # type: ignore
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate(
+                "steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False
+            )
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:  # type: ignore
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+            )
+            deprecate(
+                "clip_sample not set", "1.0.0", deprecation_message, standard_warn=False
+            )
+            new_config = dict(scheduler.config)
+            new_config["clip_sample"] = False
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            unet=unet,
+            scheduler=scheduler,
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    def enable_vae_slicing(self):
+        r"""
+        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.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError(
+                "`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher"
+            )
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError(
+                "`enable_model_offload` requires `accelerate v0.17.0` or higher."
+            )
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(
+                cpu_offloaded_model, device, prev_module_hook=hook
+            )
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.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 _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance: bool,
+        negative_prompt=None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+             prompt (`str` or `List[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 (`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. 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`).
+            prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.
+        """
+        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:
+            raise ValueError(
+                f"`prompt` should be either a string or a list of strings, but got {type(prompt)}."
+            )
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(
+            prompt, padding="longest", return_tensors="pt"
+        ).input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+            text_input_ids, untruncated_ids
+        ):
+            removed_text = self.tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+            )
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+
+        if (
+            hasattr(self.text_encoder.config, "use_attention_mask")
+            and self.text_encoder.config.use_attention_mask
+        ):
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        prompt_embeds = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        prompt_embeds = prompt_embeds[0]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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
+        )
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif 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 isinstance(negative_prompt, str):
+                uncond_tokens = [negative_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
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if (
+                hasattr(self.text_encoder.config, "use_attention_mask")
+                and self.text_encoder.config.use_attention_mask
+            ):
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(
+                dtype=self.text_encoder.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
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        return prompt_embeds
+
+    def decode_latents(self, latents):
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    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 prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            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 encode_image(self, image, device, num_images_per_prompt):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if image.dtype == np.float32:
+            image = (image * 255).astype(np.uint8)
+
+        image = self.feature_extractor(image, return_tensors="pt").pixel_values
+        image = image.to(device=device, dtype=dtype)
+
+        image_embeds = self.image_encoder(
+            image, output_hidden_states=True
+        ).hidden_states[-2]
+        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+        return torch.zeros_like(image_embeds), image_embeds
+
+    def encode_image_latents(self, image, device, num_images_per_prompt):
+
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        image = (
+            torch.from_numpy(image).unsqueeze(0).permute(0, 3, 1, 2).to(device=device)
+        )  # [1, 3, H, W]
+        image = 2 * image - 1
+        image = F.interpolate(image, (256, 256), mode="bilinear", align_corners=False)
+        image = image.to(dtype=dtype)
+
+        posterior = self.vae.encode(image).latent_dist
+        latents = posterior.sample() * self.vae.config.scaling_factor  # [B, C, H, W]
+        latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
+
+        return torch.zeros_like(latents), latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: str = "",
+        image: Optional[np.ndarray] = None,
+        height: int = 256,
+        width: int = 256,
+        elevation: float = 0,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.0,
+        negative_prompt: str = "",
+        num_images_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "numpy",  # pil, numpy, latents
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        num_frames: int = 4,
+        device=torch.device("cuda:0"),
+    ):
+        self.unet = self.unet.to(device=device)
+        self.vae = self.vae.to(device=device)
+        self.text_encoder = self.text_encoder.to(device=device)
+
+        # 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.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # imagedream variant
+        if image is not None:
+            assert isinstance(image, np.ndarray) and image.dtype == np.float32
+            self.image_encoder = self.image_encoder.to(device=device)
+            image_embeds_neg, image_embeds_pos = self.encode_image(
+                image, device, num_images_per_prompt
+            )
+            image_latents_neg, image_latents_pos = self.encode_image_latents(
+                image, device, num_images_per_prompt
+            )
+
+        _prompt_embeds = self._encode_prompt(
+            prompt=prompt,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+        )  # type: ignore
+        prompt_embeds_neg, prompt_embeds_pos = _prompt_embeds.chunk(2)
+
+        # Prepare latent variables
+        actual_num_frames = num_frames if image is None else num_frames + 1
+        latents: torch.Tensor = self.prepare_latents(
+            actual_num_frames * num_images_per_prompt,
+            4,
+            height,
+            width,
+            prompt_embeds_pos.dtype,
+            device,
+            generator,
+            None,
+        )
+
+        # Get camera
+        camera = get_camera(
+            num_frames, elevation=elevation, extra_view=(image is not None)
+        ).to(dtype=latents.dtype, device=device)
+        camera = camera.repeat_interleave(num_images_per_prompt, dim=0)
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                multiplier = 2 if do_classifier_free_guidance else 1
+                latent_model_input = torch.cat([latents] * multiplier)
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                unet_inputs = {
+                    "x": latent_model_input,
+                    "timesteps": torch.tensor(
+                        [t] * actual_num_frames * multiplier,
+                        dtype=latent_model_input.dtype,
+                        device=device,
+                    ),
+                    "context": torch.cat(
+                        [prompt_embeds_neg] * actual_num_frames
+                        + [prompt_embeds_pos] * actual_num_frames
+                    ),
+                    "num_frames": actual_num_frames,
+                    "camera": torch.cat([camera] * multiplier),
+                }
+
+                if image is not None:
+                    unet_inputs["ip"] = torch.cat(
+                        [image_embeds_neg] * actual_num_frames
+                        + [image_embeds_pos] * actual_num_frames
+                    )
+                    unet_inputs["ip_img"] = torch.cat(
+                        [image_latents_neg] + [image_latents_pos]
+                    )  # no repeat
+
+                # predict the noise residual
+                noise_pred = self.unet.forward(**unet_inputs)
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents: torch.Tensor = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                # 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:
+                        callback(i, t, latents)  # type: ignore
+
+        # Post-processing
+        if output_type == "latent":
+            image = latents
+        elif output_type == "pil":
+            image = self.decode_latents(latents)
+            image = self.numpy_to_pil(image)
+        else:  # numpy
+            image = self.decode_latents(latents)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        return image

requirements.txt

@@ -0,0 +1,55 @@
+accelerate==0.28.0
+certifi==2024.2.2
+charset-normalizer==3.3.2
+diffusers==0.27.2
+einops==0.7.0
+executing==2.0.1
+filelock==3.13.3
+fsspec==2024.3.1
+huggingface-hub==0.22.2
+idna==3.6
+importlib_metadata==7.1.0
+Jinja2==3.1.3
+kiui==0.2.7
+lazy_loader==0.3
+markdown-it-py==3.0.0
+MarkupSafe==2.1.5
+mdurl==0.1.2
+mpmath==1.3.0
+networkx==3.2.1
+numpy==1.26.4
+nvidia-cublas-cu12==12.1.3.1
+nvidia-cuda-cupti-cu12==12.1.105
+nvidia-cuda-nvrtc-cu12==12.1.105
+nvidia-cuda-runtime-cu12==12.1.105
+nvidia-cudnn-cu12==8.9.2.26
+nvidia-cufft-cu12==11.0.2.54
+nvidia-curand-cu12==10.3.2.106
+nvidia-cusolver-cu12==11.4.5.107
+nvidia-cusparse-cu12==12.1.0.106
+nvidia-nccl-cu12==2.19.3
+nvidia-nvjitlink-cu12==12.4.99
+nvidia-nvtx-cu12==12.1.105
+objprint==0.2.3
+opencv-python==4.9.0.80
+packaging==24.0
+pillow==10.2.0
+psutil==5.9.8
+Pygments==2.17.2
+PyYAML==6.0.1
+regex==2023.12.25
+requests==2.31.0
+rich==13.7.1
+safetensors==0.4.2
+scipy==1.12.0
+sympy==1.12
+tokenizers==0.15.2
+torch==2.2.2
+tqdm==4.66.2
+transformers==4.39.2
+triton==2.2.0
+typing_extensions==4.10.0
+urllib3==2.2.1
+varname==0.13.0
+xformers==0.0.25.post1
+zipp==3.18.1

scheduler/scheduler_config.json

@@ -0,0 +1,19 @@
+{
+  "_class_name": "DDIMScheduler",
+  "_diffusers_version": "0.25.0",
+  "beta_end": 0.012,
+  "beta_schedule": "scaled_linear",
+  "beta_start": 0.00085,
+  "clip_sample": false,
+  "clip_sample_range": 1.0,
+  "dynamic_thresholding_ratio": 0.995,
+  "num_train_timesteps": 1000,
+  "prediction_type": "epsilon",
+  "rescale_betas_zero_snr": false,
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": false,
+  "steps_offset": 1,
+  "thresholding": false,
+  "timestep_spacing": "leading",
+  "trained_betas": null
+}

text_encoder/config.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "stabilityai/stable-diffusion-2-1",
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1024,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 23,
+  "pad_token_id": 1,
+  "projection_dim": 512,
+  "torch_dtype": "float16",
+  "transformers_version": "4.35.2",
+  "vocab_size": 49408
+}

tokenizer/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "!",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "!",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49406": {
+      "content": "<|startoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49407": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|startoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": true,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "model_max_length": 77,
+  "pad_token": "!",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": "<|endoftext|>"
+}

unet/config.json

@@ -0,0 +1,25 @@
+{
+  "_class_name": "MultiViewUNetModel",
+  "_diffusers_version": "0.25.0",
+  "attention_resolutions": [
+    4,
+    2,
+    1
+  ],
+  "camera_dim": 16,
+  "channel_mult": [
+    1,
+    2,
+    4,
+    4
+  ],
+  "context_dim": 1024,
+  "image_size": 32,
+  "in_channels": 4,
+  "ip_dim": 16,
+  "model_channels": 320,
+  "num_head_channels": 64,
+  "num_res_blocks": 2,
+  "out_channels": 4,
+  "transformer_depth": 1
+}

unet/mv_unet.py

@@ -0,0 +1,1005 @@
+import math
+import numpy as np
+from inspect import isfunction
+from typing import Optional, Any, List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+
+from diffusers.configuration_utils import ConfigMixin
+from diffusers.models.modeling_utils import ModelMixin
+
+# require xformers!
+import xformers
+import xformers.ops
+
+from kiui.cam import orbit_camera
+
+def get_camera(
+    num_frames, elevation=15, azimuth_start=0, azimuth_span=360, blender_coord=True, extra_view=False,
+):
+    angle_gap = azimuth_span / num_frames
+    cameras = []
+    for azimuth in np.arange(azimuth_start, azimuth_span + azimuth_start, angle_gap):
+        
+        pose = orbit_camera(-elevation, azimuth, radius=1) # kiui's elevation is negated, [4, 4]
+
+        # opengl to blender
+        if blender_coord:
+            pose[2] *= -1
+            pose[[1, 2]] = pose[[2, 1]]
+
+        cameras.append(pose.flatten())
+
+    if extra_view:
+        cameras.append(np.zeros_like(cameras[0]))
+
+    return torch.from_numpy(np.stack(cameras, axis=0)).float() # [num_frames, 16]
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period)
+            * torch.arange(start=0, end=half, dtype=torch.float32)
+            / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None] * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat(
+                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
+            )
+    else:
+        embedding = repeat(timesteps, "b -> b d", d=dim)
+    # import pdb; pdb.set_trace()
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def default(val, d):
+    if val is not None:
+        return val
+    return d() if isfunction(d) else d
+
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = (
+            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
+            if not glu
+            else GEGLU(dim, inner_dim)
+        )
+
+        self.net = nn.Sequential(
+            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class MemoryEfficientCrossAttention(nn.Module):
+    # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    def __init__(
+            self, 
+            query_dim, 
+            context_dim=None, 
+            heads=8, 
+            dim_head=64, 
+            dropout=0.0,
+            ip_dim=0,
+            ip_weight=1,
+        ):
+        super().__init__()
+        
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.ip_dim = ip_dim
+        self.ip_weight = ip_weight
+
+        if self.ip_dim > 0:
+            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
+        )
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x, context=None):
+        q = self.to_q(x)
+        context = default(context, x)
+
+        if self.ip_dim > 0:
+            # context： [B, 77 + 16(ip), 1024]
+            token_len = context.shape[1]
+            context_ip = context[:, -self.ip_dim :, :]
+            k_ip = self.to_k_ip(context_ip)
+            v_ip = self.to_v_ip(context_ip)
+            context = context[:, : (token_len - self.ip_dim), :]
+
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        out = xformers.ops.memory_efficient_attention(
+            q, k, v, attn_bias=None, op=self.attention_op
+        )
+
+        if self.ip_dim > 0:
+            k_ip, v_ip = map(
+                lambda t: t.unsqueeze(3)
+                .reshape(b, t.shape[1], self.heads, self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b * self.heads, t.shape[1], self.dim_head)
+                .contiguous(),
+                (k_ip, v_ip),
+            )
+            # actually compute the attention, what we cannot get enough of
+            out_ip = xformers.ops.memory_efficient_attention(
+                q, k_ip, v_ip, attn_bias=None, op=self.attention_op
+            )
+            out = out + self.ip_weight * out_ip
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        return self.to_out(out)
+
+
+class BasicTransformerBlock3D(nn.Module):
+    
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        context_dim,
+        dropout=0.0,
+        gated_ff=True,
+        ip_dim=0,
+        ip_weight=1,
+    ):
+        super().__init__()
+
+        self.attn1 = MemoryEfficientCrossAttention(
+            query_dim=dim,
+            context_dim=None, # self-attention
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+        )
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = MemoryEfficientCrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            # ip only applies to cross-attention
+            ip_dim=ip_dim,
+            ip_weight=ip_weight,
+        ) 
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+
+    def forward(self, x, context=None, num_frames=1):
+        x = rearrange(x, "(b f) l c -> b (f l) c", f=num_frames).contiguous()
+        x = self.attn1(self.norm1(x), context=None) + x
+        x = rearrange(x, "b (f l) c -> (b f) l c", f=num_frames).contiguous()
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        context_dim, # cross attention input dim
+        depth=1,
+        dropout=0.0,
+        ip_dim=0,
+        ip_weight=1,
+    ):
+        super().__init__()
+
+        if not isinstance(context_dim, list):
+            context_dim = [context_dim]
+
+        self.in_channels = in_channels
+
+        inner_dim = n_heads * d_head
+        self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock3D(
+                    inner_dim,
+                    n_heads,
+                    d_head,
+                    context_dim=context_dim[d],
+                    dropout=dropout,
+                    ip_dim=ip_dim,
+                    ip_weight=ip_weight,
+                )
+                for d in range(depth)
+            ]
+        )
+        
+        self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
+        
+
+    def forward(self, x, context=None, num_frames=1):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context]
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = rearrange(x, "b c h w -> b (h w) c").contiguous()
+        x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context[i], num_frames=num_frames)
+        x = self.proj_out(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous()
+        
+        return x + x_in
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q, k, v = map(
+            lambda t: t.reshape(b, t.shape[1], self.heads, -1)
+            .transpose(1, 2)
+            .reshape(b, self.heads, t.shape[1], -1)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+    ):
+        super().__init__()
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim ** 0.5)
+        self.proj_in = nn.Linear(embedding_dim, dim)
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        nn.Sequential(
+                            nn.LayerNorm(dim),
+                            nn.Linear(dim, dim * ff_mult, bias=False),
+                            nn.GELU(),
+                            nn.Linear(dim * ff_mult, dim, bias=False),
+                        )
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1)
+        x = self.proj_in(x)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+class CondSequential(nn.Sequential):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, num_frames=1):
+        for layer in self:
+            if isinstance(layer, ResBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer3D):
+                x = layer(x, context, num_frames=num_frames)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(
+                dims, self.channels, self.out_channels, 3, padding=padding
+            )
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims,
+                self.channels,
+                self.out_channels,
+                3,
+                stride=stride,
+                padding=padding,
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(nn.Module):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            nn.GroupNorm(32, channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            nn.GroupNorm(32, self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class MultiViewUNetModel(ModelMixin, ConfigMixin):
+    """
+    The full multi-view UNet model with attention, timestep embedding and camera embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    :param camera_dim: dimensionality of camera input.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        transformer_depth=1,
+        context_dim=None,
+        n_embed=None,
+        num_attention_blocks=None,
+        adm_in_channels=None,
+        camera_dim=None,
+        ip_dim=0, # imagedream uses ip_dim > 0
+        ip_weight=1.0,
+        **kwargs,
+    ):
+        super().__init__()
+        assert context_dim is not None
+        
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert (
+                num_head_channels != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        if num_head_channels == -1:
+            assert (
+                num_heads != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError(
+                    "provide num_res_blocks either as an int (globally constant) or "
+                    "as a list/tuple (per-level) with the same length as channel_mult"
+                )
+            self.num_res_blocks = num_res_blocks
+        
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(
+                map(
+                    lambda i: self.num_res_blocks[i] >= num_attention_blocks[i],
+                    range(len(num_attention_blocks)),
+                )
+            )
+            print(
+                f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                f"attention will still not be set."
+            )
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        self.ip_dim = ip_dim
+        self.ip_weight = ip_weight
+
+        if self.ip_dim > 0:
+            self.image_embed = Resampler(
+                dim=context_dim,
+                depth=4,
+                dim_head=64,
+                heads=12,
+                num_queries=ip_dim,  # num token
+                embedding_dim=1280,
+                output_dim=context_dim,
+                ff_mult=4,
+            )
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            nn.Linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim),
+        )
+
+        if camera_dim is not None:
+            time_embed_dim = model_channels * 4
+            self.camera_embed = nn.Sequential(
+                nn.Linear(camera_dim, time_embed_dim),
+                nn.SiLU(),
+                nn.Linear(time_embed_dim, time_embed_dim),
+            )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(self.num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                # print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        nn.Linear(adm_in_channels, time_embed_dim),
+                        nn.SiLU(),
+                        nn.Linear(time_embed_dim, time_embed_dim),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                CondSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers: List[Any] = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+
+                    if num_attention_blocks is None or nr < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer3D(
+                                ch,
+                                num_heads,
+                                dim_head,
+                                context_dim=context_dim,
+                                depth=transformer_depth,
+                                ip_dim=self.ip_dim,
+                                ip_weight=self.ip_weight,
+                            )
+                        )
+                self.input_blocks.append(CondSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    CondSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        
+        self.middle_block = CondSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            SpatialTransformer3D(
+                ch,
+                num_heads,
+                dim_head,
+                context_dim=context_dim,
+                depth=transformer_depth,
+                ip_dim=self.ip_dim,
+                ip_weight=self.ip_weight,
+            ), 
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+
+                    if num_attention_blocks is None or i < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer3D(
+                                ch,
+                                num_heads,
+                                dim_head,
+                                context_dim=context_dim,
+                                depth=transformer_depth,
+                                ip_dim=self.ip_dim,
+                                ip_weight=self.ip_weight,
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(CondSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            nn.GroupNorm(32, ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+                nn.GroupNorm(32, ch),
+                conv_nd(dims, model_channels, n_embed, 1),
+                # nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+            )
+
+    def forward(
+        self,
+        x,
+        timesteps=None,
+        context=None,
+        y=None,
+        camera=None,
+        num_frames=1,
+        ip=None,
+        ip_img=None,
+        **kwargs,
+    ):
+        """
+        Apply the model to an input batch.
+        :param x: an [(N x F) x C x ...] Tensor of inputs. F is the number of frames (views).
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :param num_frames: a integer indicating number of frames for tensor reshaping.
+        :return: an [(N x F) x C x ...] Tensor of outputs. F is the number of frames (views).
+        """
+        assert (
+            x.shape[0] % num_frames == 0
+        ), "input batch size must be dividable by num_frames!"
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+
+        hs = []
+
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y is not None
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        # Add camera embeddings
+        if camera is not None:
+            emb = emb + self.camera_embed(camera)
+        
+        # imagedream variant
+        if self.ip_dim > 0:
+            x[(num_frames - 1) :: num_frames, :, :, :] = ip_img # place at [4, 9]
+            ip_emb = self.image_embed(ip)
+            context = torch.cat((context, ip_emb), 1)
+
+        h = x
+        for module in self.input_blocks:
+            h = module(h, emb, context, num_frames=num_frames)
+            hs.append(h)
+        h = self.middle_block(h, emb, context, num_frames=num_frames)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context, num_frames=num_frames)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

vae/config.json

@@ -0,0 +1,31 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.25.0",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 3,
+  "latent_channels": 4,
+  "layers_per_block": 2,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 256,
+  "scaling_factor": 0.18215,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ]
+}