DTG-Restore: Training-free diffusion refinement sharpens distorted videos

DTG-Restore tackles a core challenge in video super-resolution: leveraging powerful video diffusion models without costly retraining. Standard classifier-free guidance tightly couples conditional and unconditional branches, causing the model to replicate warped content from low-quality inputs. The key innovation is Decoupled Time Guidance (DTG), which evaluates the unconditional branch at a cleaner (earlier) diffusion timestep. This provides a 'lookahead' prior that guides the model toward correct geometry early in sampling, then smoothly anneals that temporal bias to let the model focus on detail refinement later—all without any fine-tuning.

The framework is plug-and-play: it works with any off-the-shelf restoration module, making it practical for real-world deployment. To stress-test the approach, the authors created GenWarp480—a benchmark of 4,400 distorted 480p videos generated from diverse text-to-video models. The benchmark targets characteristic generative degradations: warped faces, body misalignments, and spatial artifacts. Experiments show DTG-Restore significantly boosts structural fidelity (e.g., facial feature alignment) and temporal stability, outperforming prior methods that require training or specialized data. This opens the door for high-quality restoration of AI-generated and real-world video content without the compute cost of training new models.