SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
QUESTION — How can we transform low-resolution video model outputs into 4K resolution using a single-step denoising refiner?
High-resolution video generation is costly due to spatiotemporal token scaling, and multi-step refiners introduce severe sampling bottlenecks. The authors present SoL-Refiner, a single-step video refiner that converts low-resolution outputs into 4K videos through a single denoising step using a three-stage recipe of continual training, RL post-training, and one-step distillation. Evaluated on a new benchmark called Refiner-Bench, SoL-Refiner outperforms existing multi-step baselines while achieving massive latency speedups via its complete acceleration stack.
SoL-Refiner achieves an 8.91times speedup in refinement latency over the same baseline in our 2K latency setting.
Owen777 · 29 Sept 2026
read the original ↗