wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-15 16:50:57 +08:00
Code Issues Actions 12 Packages Projects Releases Wiki Activity

Rever RoPE scaling to simpler one

Browse Source
This commit is contained in:
kijai 2025-11-27 17:53:37 +02:00
parent e74db2404f
commit f1a5f6f5b3
3 changed files with 11 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
comfy/ldm/kandinsky5/model.py
View File

@ -307,23 +307,14 @@ class Kandinsky5(nn.Module):
h_start += rope_options.get("shift_y", 0.0)
w_start += rope_options.get("shift_x", 0.0)
else:
if self.model_dim == 4096: # pro video model,this is experimental as the original code only had two fixed scales for 512p and 1024p
spatial_size = h * w
scale_16384 = (1.0, 3.16, 3.16)
scale_9216 = (1.0, 2.0, 2.0)
if spatial_size <= 6144:
rope_scale_factor = scale_9216
elif spatial_size >= 14080:
rope_scale_factor = scale_16384
else:
t = (spatial_size - 14080) / (6144 - 14080)
rope_scale_factor = tuple(a + (b - a) * t for a, b in zip(scale_16384, scale_9216))
else:
rope_scale_factor = self.rope_scale_factor
rope_scale_factor = self.rope_scale_factor
if self.model_dim == 4096: # pro video model uses different rope scaling at higher resolutions
if h * w >= 14080:
rope_scale_factor = (1.0, 3.16, 3.16)
t_len = (t_len - 1.0) // rope_scale_factor[0] + 1.0
h_len = (h_len - 1.0) // rope_scale_factor[1] + 1.0
w_len = (w_len - 1.0) // rope_scale_factor[2] + 1.0
t_len = (t_len - 1.0) / rope_scale_factor[0] + 1.0
h_len = (h_len - 1.0) / rope_scale_factor[1] + 1.0
w_len = (w_len - 1.0) / rope_scale_factor[2] + 1.0
img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)

2
comfy/model_base.py
View File

@ -1690,6 +1690,6 @@ class Kandinsky5_image(Kandinsky5):
def concat_cond(self, **kwargs):
return None
def process_latent_out(self, latent): # input is still 5D, return single frame to decode with Flux VAE
return self.latent_format.process_out(latent)[:, :, 0]

6
comfy_extras/nodes_kandinsky5.py
View File

@ -67,7 +67,7 @@ def adaptive_mean_std_normalization(source, reference):
# normalization
normalized = (source - source_mean) / (source_std + 1e-8)
normalized = normalized * reference_std + reference_mean
return normalized
@ -97,9 +97,9 @@ class NormalizeVideoLatentFrames(io.ComfyNode):
first_frames = samples[:, :, :frames_to_normalize]
reference_frames_data = samples[:, :, frames_to_normalize:frames_to_normalize+min(reference_frames, samples.shape[2]-frames_to_normalize)]
normalized_first_frames = adaptive_mean_std_normalization(first_frames, reference_frames_data)
samples[:, :, :frames_to_normalize] = normalized_first_frames
s["samples"] = samples
return io.NodeOutput(s)