Merge cc337019ce into dd86b15521

comfy/text_encoders/anima.py

@@ -8,7 +8,7 @@ import torch
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):

comfy/text_encoders/flux.py

@@ -118,7 +118,7 @@ class MistralTokenizerClass:
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}
@@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class KleinTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):

comfy/text_encoders/z_image.py

@@ -6,7 +6,7 @@ import os
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 
 class ZImageTokenizer(sd1_clip.SD1Tokenizer):

comfy_extras/nodes_wan.py

@@ -27,6 +27,7 @@ class WanImageToVideo(io.ComfyNode):
                 io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
                 io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Int.Input("vae_tile_size", default=0, min=0, optional=True, tooltip="VAE encode tile size, 0 means untiled (default)"),
                 io.ClipVisionOutput.Input("clip_vision_output", optional=True),
                 io.Image.Input("start_image", optional=True),
             ],
@@ -38,14 +39,18 @@ class WanImageToVideo(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput:
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, vae_tile_size=0, start_image=None, clip_vision_output=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         if start_image is not None:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
             image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5
             image[:start_image.shape[0]] = start_image
 
-            concat_latent_image = vae.encode(image[:, :, :, :3])
+            if vae_tile_size == 0:
+                concat_latent_image = vae.encode(image[:, :, :, :3])
+            else:
+                concat_latent_image = vae.encode_tiled(image[:, :, :, :3], tile_x=vae_tile_size, tile_y=vae_tile_size, overlap=32, tile_t=256, overlap_t=8)
+
             mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype)
             mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0
 
@@ -193,6 +198,7 @@ class WanFirstLastFrameToVideo(io.ComfyNode):
                 io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
                 io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Int.Input("vae_tile_size", default=0, min=0, optional=True, tooltip="VAE encode tile size, 0 means untiled (default)"),
                 io.ClipVisionOutput.Input("clip_vision_start_image", optional=True),
                 io.ClipVisionOutput.Input("clip_vision_end_image", optional=True),
                 io.Image.Input("start_image", optional=True),
@@ -206,7 +212,7 @@ class WanFirstLastFrameToVideo(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, vae_tile_size=0, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
         spacial_scale = vae.spacial_compression_encode()
         latent = torch.zeros([batch_size, vae.latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
         if start_image is not None:
@@ -225,7 +231,11 @@ class WanFirstLastFrameToVideo(io.ComfyNode):
             image[-end_image.shape[0]:] = end_image
             mask[:, :, -end_image.shape[0]:] = 0.0
 
-        concat_latent_image = vae.encode(image[:, :, :, :3])
+        if vae_tile_size == 0:
+            concat_latent_image = vae.encode(image[:, :, :, :3])
+        else:
+            concat_latent_image = vae.encode_tiled(image[:, :, :, :3], tile_x=vae_tile_size, tile_y=vae_tile_size, overlap=32, tile_t=256, overlap_t=8)
+
         mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
         positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
         negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})