knightyxp/VideoCoF

Employs a "See → Reason → Edit" pipeline where a temporal reasoner predicts spatial reasoning tokens before generating edited video frames, enabling accurate multi-task editing (removal, addition, swap, style transfer) with just 50k training pairs. Achieves 16× length extrapolation on single-shot videos (512 frames) and 14× on multi-shot sequences despite training on only 33-frame clips, using FlashAttention-3 optimization and DMD LoRA for ~10s inference on H100 GPUs. Built on the Wan-2.1-T2V-14B foundation model with modular LoRA acceleration adapters available on Hugging Face.

159 stars.