CogVideoX 1.5

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

@@ -29,16 +29,33 @@ 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 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`.
-
-There are two models that support pose controllable generation (by the [Alibaba-PAI](https://huggingface.co/alibaba-pai) team):
-- [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose): The recommended dtype for running this model is `bf16`.
-- [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose): The recommended dtype for running this model is `bf16`.
+There are three official CogVideoX checkpoints for text-to-video and video-to-video.
+| checkpoints | recommended inference dtype |
+|---|---|
+| [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b) | torch.float16 |
+| [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b) | torch.bfloat16 |
+| [`THUDM/CogVideoX1.5-5b`](https://huggingface.co/THUDM/CogVideoX1.5-5b) | torch.bfloat16 |
+
+There are two official CogVideoX checkpoints available for image-to-video.
+| checkpoints | recommended inference dtype |
+|---|---|
+| [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V) | torch.bfloat16 |
+| [`THUDM/CogVideoX-1.5-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-1.5-5b-I2V) | torch.bfloat16 |
+- [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V): The recommended dtype for running this model is `torch.bfloat16`.
+- [`THUDM/CogVideoX1.5-5b-I2V`](https://huggingface.co/THUDM/CogVideoX1.5-5b-I2V): The recommended dtype for running this mdoel is `torch.bfloat16`.
+
+For the CogVideoX 1.5 series:
+- Text-to-video (T2V) works best at a resolution of 1360x768 because it was trained with that specific resolution.
+- Image-to-video (I2V) works for multiple resolutions. The width can vary from 768 to 1360, but the height must be 768. The height/width must be divisible by 16.
+- Both T2V and I2V models support generation with 81 and 161 frames and work best at this value. Exporting videos at 16 FPS is recommended.
+
+There are two official CogVideoX checkpoints that support pose controllable generation (by the [Alibaba-PAI](https://huggingface.co/alibaba-pai) team).
+| checkpoints | recommended inference dtype |
+|---|---|
+| [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose) | torch.bfloat16 |
+| [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose) | torch.bfloat16 |
+- [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose): The recommended dtype for running this model is `torch.bfloat16`.
+- [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose): The recommended dtype for running this model is `torch.bfloat16`.
 
 ## Inference
 

scripts/convert_cogvideox_to_diffusers.py

@@ -80,6 +80,8 @@ def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
     "post_attn1_layernorm": "norm2.norm",
     "time_embed.0": "time_embedding.linear_1",
     "time_embed.2": "time_embedding.linear_2",
+    "ofs_embed.0": "ofs_embedding.linear_1",
+    "ofs_embed.2": "ofs_embedding.linear_2",
     "mixins.patch_embed": "patch_embed",
     "mixins.final_layer.norm_final": "norm_out.norm",
     "mixins.final_layer.linear": "proj_out",
@@ -140,6 +142,7 @@ def convert_transformer(
     use_rotary_positional_embeddings: bool,
     i2v: bool,
     dtype: torch.dtype,
+    init_kwargs: Dict[str, Any],
 ):
     PREFIX_KEY = "model.diffusion_model."
 
@@ -149,7 +152,9 @@ def convert_transformer(
         num_layers=num_layers,
         num_attention_heads=num_attention_heads,
         use_rotary_positional_embeddings=use_rotary_positional_embeddings,
-        use_learned_positional_embeddings=i2v,
+        ofs_embed_dim=512 if (i2v and init_kwargs["patch_size_t"] is not None) else None,  # CogVideoX1.5-5B-I2V
+        use_learned_positional_embeddings=i2v and init_kwargs["patch_size_t"] is None,  # CogVideoX-5B-I2V
+        **init_kwargs,
     ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
@@ -163,13 +168,18 @@ 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
 
 
-def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
+def convert_vae(ckpt_path: str, scaling_factor: float, version: str, dtype: torch.dtype):
+    init_kwargs = {"scaling_factor": scaling_factor}
+    if version == "1.5":
+        init_kwargs.update({"invert_scale_latents": True})
+
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
+    vae = AutoencoderKLCogVideoX(**init_kwargs).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -187,6 +197,34 @@ def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
     return vae
 
 
+def get_transformer_init_kwargs(version: str):
+    if version == "1.0":
+        vae_scale_factor_spatial = 8
+        init_kwargs = {
+            "patch_size": 2,
+            "patch_size_t": None,
+            "patch_bias": True,
+            "sample_height": 480 // vae_scale_factor_spatial,
+            "sample_width": 720 // vae_scale_factor_spatial,
+            "sample_frames": 49,
+        }
+
+    elif version == "1.5":
+        vae_scale_factor_spatial = 8
+        init_kwargs = {
+            "patch_size": 2,
+            "patch_size_t": 2,
+            "patch_bias": False,
+            "sample_height": 300,
+            "sample_width": 300,
+            "sample_frames": 81,
+        }
+    else:
+        raise ValueError("Unsupported version of CogVideoX.")
+
+    return init_kwargs
+
+
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -202,6 +240,12 @@ def get_args():
     parser.add_argument(
         "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
     )
+    parser.add_argument(
+        "--typecast_text_encoder",
+        action="store_true",
+        default=False,
+        help="Whether or not to apply fp16/bf16 precision to text_encoder",
+    )
     # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
     parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
     # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
@@ -214,7 +258,18 @@ 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")
+    parser.add_argument(
+        "--i2v",
+        action="store_true",
+        default=False,
+        help="Whether the model to be converted is the Image-to-Video version of CogVideoX.",
+    )
+    parser.add_argument(
+        "--version",
+        choices=["1.0", "1.5"],
+        default="1.0",
+        help="Which version of CogVideoX to use for initializing default modeling parameters.",
+    )
     return parser.parse_args()
 
 
@@ -230,21 +285,27 @@ def get_args():
     dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
 
     if args.transformer_ckpt_path is not None:
+        init_kwargs = get_transformer_init_kwargs(args.version)
         transformer = convert_transformer(
             args.transformer_ckpt_path,
             args.num_layers,
             args.num_attention_heads,
             args.use_rotary_positional_embeddings,
             args.i2v,
             dtype,
+            init_kwargs,
         )
     if args.vae_ckpt_path is not None:
-        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype)
+        # Keep VAE in float32 for better quality
+        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, args.version, torch.float32)
 
     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)
 
+    if args.typecast_text_encoder:
+        text_encoder = text_encoder.to(dtype=dtype)
+
     # Apparently, the conversion does not work anymore without this :shrug:
     for param in text_encoder.parameters():
         param.data = param.data.contiguous()
@@ -276,11 +337,6 @@ def get_args():
         scheduler=scheduler,
     )
 
-    if args.fp16:
-        pipe = pipe.to(dtype=torch.float16)
-    if args.bf16:
-        pipe = pipe.to(dtype=torch.bfloat16)
-
     # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
     # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
     # is either fp16/bf16 here).

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -1057,6 +1057,7 @@ def __init__(
         force_upcast: float = True,
         use_quant_conv: bool = False,
         use_post_quant_conv: bool = False,
+        invert_scale_latents: bool = False,
     ):
         super().__init__()
 

src/diffusers/models/embeddings.py

@@ -338,6 +338,7 @@ class CogVideoXPatchEmbed(nn.Module):
     def __init__(
         self,
         patch_size: int = 2,
+        patch_size_t: Optional[int] = None,
         in_channels: int = 16,
         embed_dim: int = 1920,
         text_embed_dim: int = 4096,
@@ -355,6 +356,7 @@ def __init__(
         super().__init__()
 
         self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
         self.embed_dim = embed_dim
         self.sample_height = sample_height
         self.sample_width = sample_width
@@ -366,9 +368,15 @@ def __init__(
         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
-        )
+        if patch_size_t is None:
+            # CogVideoX 1.0 checkpoints
+            self.proj = nn.Conv2d(
+                in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
+            )
+        else:
+            # CogVideoX 1.5 checkpoints
+            self.proj = nn.Linear(in_channels * patch_size * patch_size * patch_size_t, embed_dim)
+
         self.text_proj = nn.Linear(text_embed_dim, embed_dim)
 
         if use_positional_embeddings or use_learned_positional_embeddings:
@@ -407,12 +415,24 @@ def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
         """
         text_embeds = self.text_proj(text_embeds)
 
-        batch, num_frames, channels, height, width = image_embeds.shape
-        image_embeds = image_embeds.reshape(-1, channels, height, width)
-        image_embeds = self.proj(image_embeds)
-        image_embeds = image_embeds.view(batch, num_frames, *image_embeds.shape[1:])
-        image_embeds = image_embeds.flatten(3).transpose(2, 3)  # [batch, num_frames, height x width, channels]
-        image_embeds = image_embeds.flatten(1, 2)  # [batch, num_frames x height x width, channels]
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)  # [batch, num_frames, height x width, channels]
+            image_embeds = image_embeds.flatten(1, 2)  # [batch, num_frames x height x width, channels]
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds)
 
         embeds = torch.cat(
             [text_embeds, image_embeds], dim=1
@@ -497,7 +517,14 @@ def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tens
 
 
 def get_3d_rotary_pos_embed(
-    embed_dim, crops_coords, grid_size, temporal_size, theta: int = 10000, use_real: bool = True
+    embed_dim,
+    crops_coords,
+    grid_size,
+    temporal_size,
+    theta: int = 10000,
+    use_real: bool = True,
+    grid_type: str = "linspace",
+    max_size: Optional[Tuple[int, int]] = None,
 ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
     """
     RoPE for video tokens with 3D structure.
@@ -513,17 +540,30 @@ def get_3d_rotary_pos_embed(
         The size of the temporal dimension.
     theta (`float`):
         Scaling factor for frequency computation.
+    grid_type (`str`):
+        Whether to use "linspace" or "slice" to compute grids.
 
     Returns:
         `torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
     """
     if use_real is not True:
         raise ValueError(" `use_real = False` is not currently supported for get_3d_rotary_pos_embed")
-    start, stop = crops_coords
-    grid_size_h, grid_size_w = grid_size
-    grid_h = np.linspace(start[0], stop[0], grid_size_h, endpoint=False, dtype=np.float32)
-    grid_w = np.linspace(start[1], stop[1], grid_size_w, endpoint=False, dtype=np.float32)
-    grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
+
+    if grid_type == "linspace":
+        start, stop = crops_coords
+        grid_size_h, grid_size_w = grid_size
+        grid_h = np.linspace(start[0], stop[0], grid_size_h, endpoint=False, dtype=np.float32)
+        grid_w = np.linspace(start[1], stop[1], grid_size_w, endpoint=False, dtype=np.float32)
+        grid_t = np.arange(temporal_size, dtype=np.float32)
+        grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
+    elif grid_type == "slice":
+        max_h, max_w = max_size
+        grid_size_h, grid_size_w = grid_size
+        grid_h = np.arange(max_h, dtype=np.float32)
+        grid_w = np.arange(max_w, dtype=np.float32)
+        grid_t = np.arange(temporal_size, dtype=np.float32)
+    else:
+        raise ValueError("Invalid value passed for `grid_type`.")
 
     # Compute dimensions for each axis
     dim_t = embed_dim // 4
@@ -559,6 +599,12 @@ def combine_time_height_width(freqs_t, freqs_h, freqs_w):
     t_cos, t_sin = freqs_t  # both t_cos and t_sin has shape: temporal_size, dim_t
     h_cos, h_sin = freqs_h  # both h_cos and h_sin has shape: grid_size_h, dim_h
     w_cos, w_sin = freqs_w  # both w_cos and w_sin has shape: grid_size_w, dim_w
+
+    if grid_type == "slice":
+        t_cos, t_sin = t_cos[:temporal_size], t_sin[:temporal_size]
+        h_cos, h_sin = h_cos[:grid_size_h], h_sin[:grid_size_h]
+        w_cos, w_sin = w_cos[:grid_size_w], w_sin[:grid_size_w]
+
     cos = combine_time_height_width(t_cos, h_cos, w_cos)
     sin = combine_time_height_width(t_sin, h_sin, w_sin)
     return cos, sin