SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
The paper claims a one-step refiner can turn low-resolution video outputs into 4K video while avoiding the usual second sampling bottleneck.
SoL-Refiner uses high-resolution continual training, RL post-training, and one-step distillation. The authors introduce Refiner-Bench to compare refiners on shared low-resolution inputs from different video generators. In their tests, SoL-Refiner beats external refiners on VBench and UniPercept averages at about 2K, and improves those metrics over the three-step LTX-2.3 Refiner at 3840×2176. With its full acceleration stack, it reports an 8.91× refinement-latency speedup over that baseline in the 2K setting. HF Daily Papers' note
SoL-Refiner uses high-resolution continual training, RL post-training, and one-step distillation. The authors introduce Refiner-Bench to compare refiners on shared low-resolution inputs from different video generators. In their tests, SoL-Refiner beats external refiners on VBench and UniPercept averages at about 2K, and improves those metrics over the three-step LTX-2.3 Refiner at 3840×2176. With its full acceleration stack, it reports an 8.91× refinement-latency speedup over that baseline in the 2K setting. HF Daily Papers' note
score 5