[pull] master from Comfy-Org:master

comfy/latent_formats.py

@@ -1,4 +1,5 @@
 import torch
+import comfy.nested_tensor
 
 class LatentFormat:
     scale_factor = 1.0
@@ -17,6 +18,9 @@ def process_in(self, latent):
     def process_out(self, latent):
         return latent / self.scale_factor
 
+    def fix_empty_latent(self, latent):
+        return latent
+
 class SD15(LatentFormat):
     def __init__(self, scale_factor=0.18215):
         self.scale_factor = scale_factor
@@ -606,6 +610,19 @@ class MiniMaxH3AV(MiniMaxH3Video):
     # max channels across the two streams (video 24, audio 32) so per-stream slices keep both streams whole
     latent_channels = 32
 
+    def fix_empty_latent(self, latent):
+        video_latent_channels = MiniMaxH3Video.latent_channels
+        audio_latent_channels = 32
+        audio_channels = 2
+        frames_per_token = (1, 4, 4, 4, 4)
+        audio_frame_rescale = 5.0 / 3.0
+
+        video = latent[:, :video_latent_channels].clone()
+        frame_count = sum(frames_per_token[i % len(frames_per_token)] for i in range(video.shape[2]))
+        audio_t = round(frame_count * audio_frame_rescale)
+        audio = latent.new_zeros((latent.shape[0], audio_latent_channels, audio_channels, audio_t))
+        return comfy.nested_tensor.NestedTensor((video, audio))
+
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3

comfy/sample.py

@@ -45,7 +45,7 @@ def prepare_empty_noise(latent_image):
 def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None, downscale_ratio_temporal=None):
     if latent_image.is_nested:
         return latent_image
-    latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels
+    latent_format = model.get_model_object("latent_format")
     is_empty = torch.count_nonzero(latent_image) == 0
     if is_empty:
         if latent_format.latent_channels != latent_image.shape[1]:
@@ -64,6 +64,9 @@ def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None,
             new_t = max(1, round(latent_image.shape[2] * ratio))
             latent_image = comfy.utils.repeat_to_batch_size(latent_image, new_t, dim=2)
 
+    if is_empty:
+        latent_image = latent_format.fix_empty_latent(latent_image)
+
     return latent_image
 
 def prepare_sampling(model, noise_shape, positive, negative, noise_mask):