G-U-N/Gen-L-Video

Extends existing short-video diffusion models to generate and edit videos with hundreds of frames and multiple semantic segments through temporal co-denoising, without additional training. Supports diverse conditioning modalities (pose, depth, sketch, canny, semantic segmentation) via T2I-Adapters and ControlNet, plus spatial editing through SAM and Grounding DINO integration. Offers both one-shot tuning and training-free inference pipelines compatible with Stable Diffusion-based architectures.

307 stars.