Ethanfel/Comfyui-STAR
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Move VAE encode outside autocast to match original STAR pipeline

Browse Source 
The original STAR code runs vae_encode() before the amp.autocast() block.
Our code had it inside, which changes how the encoder processes tensors
and can produce different latent representations.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Ethanfel
2026-02-15 02:14:25 +01:00
parent 0537d9d8a5
commit e2025c6ca0
1 changed files with 11 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
star_pipeline.py
View File

@@ -302,18 +302,18 @@ def run_star_inference(
text_encoder.device = "cpu"
torch.cuda.empty_cache()
# -- Diffusion sampling (autocast needed for fp16 VAE / UNet) --
with torch.amp.autocast("cuda"):
# ---- Stage 2: VAE encode ----
if offload != "disabled":
_move(vae, device)
video_data_feature = vae_encode(vae, video_data, chunk_size=vae_enc_chunk)
if offload != "disabled":
_move(vae, "cpu")
# Free the full-res pixel tensor — only latents needed from here.
del video_data
torch.cuda.empty_cache()
# ---- Stage 2: VAE encode (outside autocast, matches original STAR) ----
if offload != "disabled":
_move(vae, device)
video_data_feature = vae_encode(vae, video_data, chunk_size=vae_enc_chunk)
if offload != "disabled":
_move(vae, "cpu")
# Free the full-res pixel tensor — only latents needed from here.
del video_data
torch.cuda.empty_cache()
# -- Diffusion sampling + VAE decode (under autocast) --
with torch.amp.autocast("cuda"):
t = torch.LongTensor([total_noise_levels - 1]).to(device)
noised_lr = diffusion.diffuse(video_data_feature, t)