HorizonRelight: Relighting Long-horizon Videos Consistently via Diffusion Transformers

The blog post discusses a new framework for relighting long-horizon videos using diffusion transformers, addressing challenges in temporal consistency during chunked video processing. By reframing the problem as temporally conditioned latent domain translation and introducing techniques like masked target-domain self-conditioning and warm-start prompting, the approach aims to enhance video quality across chunk boundaries, resulting in improved continuity and reduced artifacts in generated content.