17 lines
1.1 KiB
Plaintext
17 lines
1.1 KiB
Plaintext
tree 280b4fbb309f97ac668e0ed528a12b98f1a12e17
|
||
parent 8456308f448f52591c949f6e2000c6b6478ca70e
|
||
author wuwukasi <wuwukawayi@gmail.com> 1779454024 +0800
|
||
committer wuwukasi <wuwukawayi@gmail.com> 1779454024 +0800
|
||
|
||
feat: unified canvas expansion for prefix and transition video
|
||
|
||
- Add dynamic prefix frames (1-5 images, img0×1 + imgN×4 pattern)
|
||
- Expand canvas by 21/37 frames at encode time (transition/prefix)
|
||
- Embed prefix and transition into canvas before VAE encode, removing per-window injection logic from the sampling loop
|
||
- Unify pose/face/mask/bg padding to stride-2 reversed sampling
|
||
- Support looping mode: prefix_ctx (ref+prefix latent) prepended to every chunk like ref_latent; pose/face copy first frame of current chunk to fill prefix region
|
||
- Dynamic decode trim via canvas_expansion_px (21/37/0)
|
||
- Fix latent_window_size to cover full expanded bg range, preventing pose index clamping in non-first context windows
|
||
- Pad uni3c render_latent with 10 leading first-frame copies when prefix is active
|
||
- Remove independent VAE encode for transition (now embedded in canvas)
|