Void model - pass 1 & 2 (CORE-38) (Comfy-Org#13403)

Author: Talmaj

comfy/latent_formats.py

@@ -793,9 +793,27 @@ class ZImagePixelSpace(ChromaRadiance):
     pass
 
 class CogVideoX(LatentFormat):
+    """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b).
+
+    scale_factor matches the vae/config.json scaling_factor for the 2b variant.
+    The 5b-class checkpoints (CogVideoX-5b, CogVideoX-1.5-5B, CogVideoX-Fun-V1.5-*)
+    use a different value; see CogVideoX1_5 below.
+    """
     latent_channels = 16
     latent_dimensions = 3
     temporal_downscale_ratio = 4
 
     def __init__(self):
         self.scale_factor = 1.15258426
+
+
+class CogVideoX1_5(CogVideoX):
+    """Latent format for 5b-class CogVideoX checkpoints.
+
+    Covers THUDM/CogVideoX-5b, THUDM/CogVideoX-1.5-5B, and the CogVideoX-Fun
+    V1.5-5b family (including VOID inpainting). All of these have
+    scaling_factor=0.7 in their vae/config.json. Auto-selected in
+    supported_models.CogVideoX_T2V based on transformer hidden dim.
+    """
+    def __init__(self):
+        self.scale_factor = 0.7

comfy/sd.py

@@ -66,6 +66,7 @@
 import comfy.text_encoders.qwen35
 import comfy.text_encoders.ernie
 import comfy.text_encoders.gemma4
+import comfy.text_encoders.cogvideo
 
 import comfy.model_patcher
 import comfy.lora
@@ -1224,6 +1225,7 @@ class CLIPType(Enum):
     NEWBIE = 24
     FLUX2 = 25
     LONGCAT_IMAGE = 26
+    COGVIDEOX = 27
 
 
 
@@ -1428,6 +1430,9 @@ class EmptyClass:
                 clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+            elif clip_type == CLIPType.COGVIDEOX:
+                clip_target.clip = comfy.text_encoders.cogvideo.cogvideo_te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.cogvideo.CogVideoXTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer

comfy/supported_models.py

@@ -1872,6 +1872,14 @@ class CogVideoX_T2V(supported_models_base.BASE):
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        # 2b-class (dim=1920, heads=30) uses scale_factor=1.15258426.
+        # 5b-class (dim=3072, heads=48) — incl. CogVideoX-5b, 1.5-5B, and
+        # Fun-V1.5 inpainting — uses scale_factor=0.7 per vae/config.json.
+        if unet_config.get("num_attention_heads", 0) >= 48:
+            self.latent_format = latent_formats.CogVideoX1_5
+        super().__init__(unet_config)
+
     def get_model(self, state_dict, prefix="", device=None):
         # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
         if self.unet_config.get("patch_size_t") is not None:
@@ -1898,6 +1906,20 @@ def get_model(self, state_dict, prefix="", device=None):
         out = model_base.CogVideoX(self, image_to_video=True, device=device)
         return out
 
+class CogVideoX_Inpaint(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
 
 models = [
     LotusD,
@@ -1978,6 +2000,7 @@ def get_model(self, state_dict, prefix="", device=None):
     ErnieImage,
     SAM3,
     SAM31,
+    CogVideoX_Inpaint,
     CogVideoX_I2V,
     CogVideoX_T2V,
     SVD_img2vid,

comfy/text_encoders/cogvideo.py

@@ -1,6 +1,48 @@
 import comfy.text_encoders.sd3_clip
+from comfy import sd1_clip
 
 
 class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    """Inner T5 tokenizer for CogVideoX.
+
+    CogVideoX was trained with T5 embeddings padded to 226 tokens (not 77 like SD3).
+    Used both directly by supported_models.CogVideoX_T2V.clip_target (paired with
+    the raw T5XXLModel) and by the CogVideoXTokenizer outer wrapper below.
+    """
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)
+
+
+class CogVideoXTokenizer(sd1_clip.SD1Tokenizer):
+    """Outer tokenizer wrapper for CLIPLoader (type="cogvideox")."""
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data,
+                         clip_name="t5xxl", tokenizer=CogVideoXT5Tokenizer)
+
+
+class CogVideoXT5XXL(sd1_clip.SD1ClipModel):
+    """Outer T5XXL model wrapper for CLIPLoader (type="cogvideox").
+
+    Wraps the raw T5XXL model in the SD1ClipModel interface so that CLIP.__init__
+    (which reads self.dtypes) works correctly. The inner model is the standard
+    sd3_clip.T5XXLModel (no attention_mask change needed for CogVideoX).
+    """
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl",
+                         clip_model=comfy.text_encoders.sd3_clip.T5XXLModel,
+                         model_options=model_options)
+
+
+def cogvideo_te(dtype_t5=None, t5_quantization_metadata=None):
+    """Factory that returns a CogVideoXT5XXL class configured with the detected
+    T5 dtype and optional quantization metadata, for use in load_text_encoder_state_dicts.
+    """
+    class CogVideoXTEModel_(CogVideoXT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return CogVideoXTEModel_