Switch apply-step from Wan-Animate to Wan2.1-VACE masked inpaint: regenerates ONLY the clothing region, copies face/hair/skin/motion/background from the source at original sharpness. UI: video + prompt -> video. VACE on 'large' (49 frames @6fps, ~265 quota incl remote FireRed) fits a FREE account. Boot probes reference_images + pack size. Removed dead wan_engine helpers; stopped purging the VACE model cache.

Adopt alexnasa/Wan2.2-Animate-ZEROGPU mechanism wholesale: official wan package, GPU pose (torch 2.8/CUDA12 + onnxruntime-gpu), clip slicing. Default mode = Pose Retarget. Root cause of our slowness was CUDA-13 forcing onnxruntime pose onto CPU (1.87s/frame).

Free-tier support: segment_frame_length lever (seg 41/5 steps) so a FREE 300s/day account fits one run (290 quota); make quota_probe a 0-cost info endpoint (was 580 quota, unusable by free accounts)

Restore Wan-VACE prompt-based clothing editor (revert Animate swap)

Replace with alexnasa/Wan2.2-Animate-ZEROGPU

FireRed keyframes: actually anchor later frames to frame 1 + denser keyframes

Switch to FireRed keyframe edit + RIFE interpolation (keeps the face)

FP8-quantize FireRed transformer so both resident models fit the disk pack

Cut FireRed GPU reservation 200->70 (was 400s xlarge, rejected as over-cap)

Revert to 1024px editing + add framing-preservation prompt (low res made Qwen zoom/recompose close-ups)

Edit at 640px + 72 keyframes: ~2x denser real-motion sampling so motion tracks the source (RIFE invents less)

Edit each frame independently (anchor approach left moving frames unchanged / outfit not replaced)

Add output-fps selector (20/30/60); RIFE interpolates to chosen fps

Add RIFE interpolation: edit ~48 keyframes then interpolate up to ~24fps (fixes choppy 6fps)

firered_edit_video: add size='xlarge' (57GB model OOM'd on default partition)

Pivot to frame-by-frame clothing edit with FireRed; drop WanAnimate (28GB model / CUDA-12/13 conflict / timeouts)

Drop FireRed full-repo prefetch (blew 150G disk); from_pretrained fetches only needed files

FireRed: disable FA3 (crashes ZeroGPU worker), use xlarge GPU size + 200s

Force YOLO detection onto CPU (was grabbing CUDA in main worker and aborting the FireRed GPU task); bump FireRed GPU reservation to 120s

Fix FireRed NVML crash: load pipeline at import (like parent Space), import engine at boot

Fix cold-start: move weight downloads + flash-attn off the GPU clock; cap Wan reservation under 5min; default 5s clip

Upload 10 files