thu-ml/DiT-Extrapolation

Implements plug-and-play length extrapolation for video diffusion transformers by modifying a single RoPE frequency parameter, enabling models like HunyuanVideo and CogVideoX to generate longer sequences (5s→11s) without retraining. Extends to high-resolution image generation through UltraImage branch and supports multiple inference backends including Diffusers, with pre-trained checkpoints available on HuggingFace. Integrates with ComfyUI and other inference frameworks, offering both single-GPU and multi-GPU inference options.

794 stars. Actively maintained with 1 commit in the last 30 days.