CogVideoX-5b-I2V support

docs/source/en/api/pipelines/cogvideox.md

@@ -29,9 +29,12 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
 
-There are two models available that can be used with the CogVideoX pipeline:
-- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b)
-- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b)
+There are two models available that can be used with the text-to-video and video-to-video CogVideoX pipelines:
+- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b): The recommended dtype for running this model is `fp16`.
+- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b): The recommended dtype for running this model is `bf16`.
+
+There is one model available that can be used with the image-to-video CogVideoX pipeline:
+- [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V): The recommended dtype for running this model is `bf16`.
 
 ## Inference
 
@@ -98,6 +101,12 @@ It is also worth noting that torchao quantization is fully compatible with [torc
   - all
   - __call__
 
+## CogVideoXImageToVideoPipeline
+
+[[autodoc]] CogVideoXImageToVideoPipeline
+  - all
+  - __call__
+
 ## CogVideoXVideoToVideoPipeline
 
 [[autodoc]] CogVideoXVideoToVideoPipeline

scripts/convert_cogvideox_to_diffusers.py

@@ -4,7 +4,13 @@
 import torch
 from transformers import T5EncoderModel, T5Tokenizer
 
-from diffusers import AutoencoderKLCogVideoX, CogVideoXDDIMScheduler, CogVideoXPipeline, CogVideoXTransformer3DModel
+from diffusers import (
+    AutoencoderKLCogVideoX,
+    CogVideoXDDIMScheduler,
+    CogVideoXImageToVideoPipeline,
+    CogVideoXPipeline,
+    CogVideoXTransformer3DModel,
+)
 
 
 def reassign_query_key_value_inplace(key: str, state_dict: Dict[str, Any]):
@@ -78,6 +84,7 @@ def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
     "mixins.final_layer.norm_final": "norm_out.norm",
     "mixins.final_layer.linear": "proj_out",
     "mixins.final_layer.adaLN_modulation.1": "norm_out.linear",
+    "mixins.pos_embed.pos_embedding": "patch_embed.pos_embedding",  # Specific to CogVideoX-5b-I2V
 }
 
 TRANSFORMER_SPECIAL_KEYS_REMAP = {
@@ -131,15 +138,18 @@ def convert_transformer(
     num_layers: int,
     num_attention_heads: int,
     use_rotary_positional_embeddings: bool,
+    i2v: bool,
     dtype: torch.dtype,
 ):
     PREFIX_KEY = "model.diffusion_model."
 
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
     transformer = CogVideoXTransformer3DModel(
+        in_channels=32 if i2v else 16,
         num_layers=num_layers,
         num_attention_heads=num_attention_heads,
         use_rotary_positional_embeddings=use_rotary_positional_embeddings,
+        use_learned_positional_embeddings=i2v,
     ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
@@ -153,7 +163,6 @@ def convert_transformer(
             if special_key not in key:
                 continue
             handler_fn_inplace(key, original_state_dict)
-
     transformer.load_state_dict(original_state_dict, strict=True)
     return transformer
 
@@ -205,6 +214,7 @@ def get_args():
     parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
     # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
     parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
+    parser.add_argument("--i2v", action="store_true", default=False, help="Whether to save the model weights in fp16")
     return parser.parse_args()
 
 
@@ -225,6 +235,7 @@ def get_args():
             args.num_layers,
             args.num_attention_heads,
             args.use_rotary_positional_embeddings,
+            args.i2v,
             dtype,
         )
     if args.vae_ckpt_path is not None:
@@ -233,7 +244,6 @@ def get_args():
     text_encoder_id = "google/t5-v1_1-xxl"
     tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
     text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
-
     # Apparently, the conversion does not work any more without this :shrug:
     for param in text_encoder.parameters():
         param.data = param.data.contiguous()
@@ -252,9 +262,17 @@ def get_args():
             "timestep_spacing": "trailing",
         }
     )
-
-    pipe = CogVideoXPipeline(
-        tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+    if args.i2v:
+        pipeline_cls = CogVideoXImageToVideoPipeline
+    else:
+        pipeline_cls = CogVideoXPipeline
+
+    pipe = pipeline_cls(
+        tokenizer=tokenizer,
+        text_encoder=text_encoder,
+        vae=vae,
+        transformer=transformer,
+        scheduler=scheduler,
     )
 
     if args.fp16:

src/diffusers/__init__.py

@@ -255,6 +255,7 @@
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
+            "CogVideoXImageToVideoPipeline",
             "CogVideoXPipeline",
             "CogVideoXVideoToVideoPipeline",
             "CycleDiffusionPipeline",
@@ -703,6 +704,7 @@
             AudioLDMPipeline,
             AuraFlowPipeline,
             CLIPImageProjection,
+            CogVideoXImageToVideoPipeline,
             CogVideoXPipeline,
             CogVideoXVideoToVideoPipeline,
             CycleDiffusionPipeline,

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -1089,8 +1089,10 @@ def _encode(self, x: torch.Tensor) -> torch.Tensor:
             return self.tiled_encode(x)
 
         frame_batch_size = self.num_sample_frames_batch_size
+        # Note: We expect the number of frames to be either `1` or `frame_batch_size * k` or `frame_batch_size * k + 1` for some k.
+        num_batches = num_frames // frame_batch_size if num_frames > 1 else 1
         enc = []
-        for i in range(num_frames // frame_batch_size):
+        for i in range(num_batches):
             remaining_frames = num_frames % frame_batch_size
             start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
             end_frame = frame_batch_size * (i + 1) + remaining_frames
@@ -1140,8 +1142,9 @@ def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOut
             return self.tiled_decode(z, return_dict=return_dict)
 
         frame_batch_size = self.num_latent_frames_batch_size
+        num_batches = num_frames // frame_batch_size
         dec = []
-        for i in range(num_frames // frame_batch_size):
+        for i in range(num_batches):
             remaining_frames = num_frames % frame_batch_size
             start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
             end_frame = frame_batch_size * (i + 1) + remaining_frames
@@ -1233,8 +1236,10 @@ def tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
         for i in range(0, height, overlap_height):
             row = []
             for j in range(0, width, overlap_width):
+                # Note: We expect the number of frames to be either `1` or `frame_batch_size * k` or `frame_batch_size * k + 1` for some k.
+                num_batches = num_frames // frame_batch_size if num_frames > 1 else 1
                 time = []
-                for k in range(num_frames // frame_batch_size):
+                for k in range(num_batches):
                     remaining_frames = num_frames % frame_batch_size
                     start_frame = frame_batch_size * k + (0 if k == 0 else remaining_frames)
                     end_frame = frame_batch_size * (k + 1) + remaining_frames
@@ -1309,8 +1314,9 @@ def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[Decod
         for i in range(0, height, overlap_height):
             row = []
             for j in range(0, width, overlap_width):
+                num_batches = num_frames // frame_batch_size
                 time = []
-                for k in range(num_frames // frame_batch_size):
+                for k in range(num_batches):
                     remaining_frames = num_frames % frame_batch_size
                     start_frame = frame_batch_size * k + (0 if k == 0 else remaining_frames)
                     end_frame = frame_batch_size * (k + 1) + remaining_frames

src/diffusers/models/embeddings.py

@@ -350,6 +350,7 @@ def __init__(
         spatial_interpolation_scale: float = 1.875,
         temporal_interpolation_scale: float = 1.0,
         use_positional_embeddings: bool = True,
+        use_learned_positional_embeddings: bool = True,
     ) -> None:
         super().__init__()
 
@@ -363,15 +364,17 @@ def __init__(
         self.spatial_interpolation_scale = spatial_interpolation_scale
         self.temporal_interpolation_scale = temporal_interpolation_scale
         self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
 
         self.proj = nn.Conv2d(
             in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
         )
         self.text_proj = nn.Linear(text_embed_dim, embed_dim)
 
-        if use_positional_embeddings:
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
             pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
-            self.register_buffer("pos_embedding", pos_embedding, persistent=False)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
 
     def _get_positional_embeddings(self, sample_height: int, sample_width: int, sample_frames: int) -> torch.Tensor:
         post_patch_height = sample_height // self.patch_size
@@ -415,8 +418,15 @@ def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
             [text_embeds, image_embeds], dim=1
         ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
 
-        if self.use_positional_embeddings:
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            if self.use_learned_positional_embeddings and (self.sample_width != width or self.sample_height != height):
+                raise ValueError(
+                    "It is currently not possible to generate videos at a different resolution that the defaults. This should only be the case with 'THUDM/CogVideoX-5b-I2V'."
+                    "If you think this is incorrect, please open an issue at https://github.com/huggingface/diffusers/issues."
+                )
+
             pre_time_compression_frames = (num_frames - 1) * self.temporal_compression_ratio + 1
+
             if (
                 self.sample_height != height
                 or self.sample_width != width

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -235,6 +235,7 @@ def __init__(
         spatial_interpolation_scale: float = 1.875,
         temporal_interpolation_scale: float = 1.0,
         use_rotary_positional_embeddings: bool = False,
+        use_learned_positional_embeddings: bool = False,
     ):
         super().__init__()
         inner_dim = num_attention_heads * attention_head_dim
@@ -254,6 +255,7 @@ def __init__(
             spatial_interpolation_scale=spatial_interpolation_scale,
             temporal_interpolation_scale=temporal_interpolation_scale,
             use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
         )
         self.embedding_dropout = nn.Dropout(dropout)
 
@@ -465,8 +467,11 @@ def custom_forward(*inputs):
         hidden_states = self.proj_out(hidden_states)
 
         # 5. Unpatchify
+        # Note: we use `-1` instead of `channels`:
+        #   - It is okay to `channels` use for CogVideoX-2b and CogVideoX-5b (number of input channels is equal to output channels)
+        #   - However, for CogVideoX-5b-I2V also takes concatenated input image latents (number of input channels is twice the output channels)
         p = self.config.patch_size
-        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
         output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
 
         if not return_dict:

src/diffusers/pipelines/__init__.py

@@ -138,7 +138,11 @@
         "AudioLDM2UNet2DConditionModel",
     ]
     _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
-    _import_structure["cogvideo"] = ["CogVideoXPipeline", "CogVideoXVideoToVideoPipeline"]
+    _import_structure["cogvideo"] = [
+        "CogVideoXPipeline",
+        "CogVideoXImageToVideoPipeline",
+        "CogVideoXVideoToVideoPipeline",
+    ]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -461,7 +465,7 @@
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
-        from .cogvideo import CogVideoXPipeline, CogVideoXVideoToVideoPipeline
+        from .cogvideo import CogVideoXImageToVideoPipeline, CogVideoXPipeline, CogVideoXVideoToVideoPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
             StableDiffusionControlNetImg2ImgPipeline,

src/diffusers/pipelines/cogvideo/__init__.py

@@ -23,6 +23,7 @@
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_cogvideox"] = ["CogVideoXPipeline"]
+    _import_structure["pipeline_cogvideox_image2video"] = ["CogVideoXImageToVideoPipeline"]
     _import_structure["pipeline_cogvideox_video2video"] = ["CogVideoXVideoToVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -34,6 +35,7 @@
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_cogvideox import CogVideoXPipeline
+        from .pipeline_cogvideox_image2video import CogVideoXImageToVideoPipeline
         from .pipeline_cogvideox_video2video import CogVideoXVideoToVideoPipeline
 
 else: