EzioBy/Ditto

Combines an image editor with in-context video generation to synthetically create Ditto-1M, a million-example dataset addressing data scarcity in instruction-based video editing. The framework uses a distilled model architecture with temporal enhancement and an intelligent agent for instruction generation and quality filtering, trained with curriculum learning on the Wan video diffusion backbone. Provides inference through DiffSynth-Studio and ComfyUI, with support for multi-node training via DLRover and post-generation denoising enhancement.

585 stars.