Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Do padding of audio embed in model for humo for more flexibility. (#9935)

comfyanonymous24b0fce099c5
diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py
index 9cf3c171d..2dac5980c 100644
--- a/comfy/ldm/wan/model.py
+++ b/comfy/ldm/wan/model.py
@@ -1551,6 +1551,9 @@ class HumoWanModel(WanModel):
context_img_len = None
if audio_embed is not None:
+ if reference_latent is not None:
+ zero_audio_pad = torch.zeros(audio_embed.shape[0], reference_latent.shape[-3], *audio_embed.shape[2:], device=audio_embed.device, dtype=audio_embed.dtype)
+ audio_embed = torch.cat([audio_embed, zero_audio_pad], dim=1)
audio = self.audio_proj(audio_embed).permute(0, 3, 1, 2).flatten(2).transpose(1, 2)
else:
audio = None
diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py
index 0b8b55813..5f10edcff 100644
--- a/comfy_extras/nodes_wan.py
+++ b/comfy_extras/nodes_wan.py
@@ -1095,10 +1095,6 @@ class WanHuMoImageToVideo(io.ComfyNode):
audio_emb = torch.stack([feat0, feat1, feat2, feat3, feat4], dim=2)[0] # [T, 5, 1280]
audio_emb, _ = get_audio_emb_window(audio_emb, length, frame0_idx=0)
- # pad for ref latent
- zero_audio_pad = torch.zeros(ref_latent.shape[2], *audio_emb.shape[1:], device=audio_emb.device, dtype=audio_emb.dtype)
- audio_emb = torch.cat([audio_emb, zero_audio_pad], dim=0)
-
audio_emb = audio_emb.unsqueeze(0)
audio_emb_neg = torch.zeros_like(audio_emb)
positive = node_helpers.conditioning_set_values(positive, {"audio_embed": audio_emb})