SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
Haozhe Liu, Tian Ye, Shuchen Xue, Yitong Li, Junsong Chen, Haopeng Li, Jincheng Yu, Duomin Wang and 4 others
NVIDIA
Published September 29, 2026 · on the feed September 30, 2026 · ▲ 39 · arXiv 2609.37969
The abstract
High-resolution video generation is expensive, as its cost grows rapidly with the number of spatiotemporal tokens. A practical alternative first generates a lower-resolution video and then applies a refiner, but conventional multi-step refinement introduces a second sampling bottleneck. We present SoL-Refiner, a one-step video refiner that transforms low-resolution model outputs into 4K videos with a single denoising step. Our three-stage recipe combines high-resolution continual training, reinforcement learning (RL) post-training, and a final one-step distillation. We introduce Refiner-Bench, a video refinement benchmark constructed from the outputs of different video generators, and use a shared-input protocol to compare refiners at approximately 2K output resolution. At 2K, the one-step SoL-Refiner outperforms all external refiners on the VBench and UniPercept averages, while at 3840!times!2176 it improves both metrics over the three-step LTX-2.3 Refiner. With the complete acceleration stack, SoL-Refiner achieves an 8.91times speedup in refinement latency over the same baseline in our 2K latency setting.
The Brief
1:50A short podcast summary of the paper. Written and voiced by AI.
Transcript
SoL-Refiner. Speed-of-Light One-Step Refinement for High-Resolution Video. By Haozhe Liu and eleven others, from NVIDIA. Published on the twenty-ninth of September. No code repository is linked to it.
Generating high-resolution video is expensive, because the cost grows rapidly with the number of spatiotemporal tokens. A practical workaround is to generate at a lower resolution first, and then hand the result to a refiner. But a conventional refiner takes several sampling steps of its own, and that brings back a second sampling bottleneck at the very stage meant to save time.
SoL-Refiner removes the bottleneck by refining in one step. It takes a low-resolution model output and turns it into a four K video with a single denoising step.
Getting there takes a three-stage recipe. First comes continual training at high resolution. Then reinforcement learning as post-training. Last, a one-step distillation. To compare refiners fairly, the authors also introduce Refiner-Bench, a video refinement benchmark built from the outputs of different video generators. Under a shared-input protocol, every refiner receives the same inputs, with output at roughly two K resolution.
At two K, the one-step SoL-Refiner beats every external refiner on both the VBench and UniPercept averages. At the full output resolution of three thousand eight hundred and forty by two thousand one hundred and seventy-six, it improves both metrics over the LTX-2.3 Refiner, which uses three steps. With the complete acceleration stack in place, refinement latency drops by a factor of eight point nine one against that same baseline, measured in the two K latency setting.
A refiner exists to save the cost of generating at full resolution, so it should not become a slow stage in its own right. Collapsing it to a single step is the direct answer, and here the single step does not give up quality to the three-step baseline it was measured against.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
High resolution is beautiful and slow. The shortcut has a slow step of its own.
You want sharp 4K video from a generator, but cost grows quickly with the number of spatiotemporal tokens. Sampling at full resolution is painfully slow.
The road ahead
- 1 · The problem ~6 min Making sharp 4K video is slow, and the usual low-resolution-then-refine workaround brings the slowness back in its refinement stage.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Train a refiner that turns a low-resolution video into high-resolution video in one denoising step, built up through three training stages.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train at high resolution, post-train with rewards, distill to one step, then refine low-resolution outputs into 4K in one pass.
- 6 · The experiment ~6 min You’ll guess: With the complete acceleration stack, how much faster was SoL-Refiner's refinement than the three-step LTX-2.3 Refiner in the 2K latency setting?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.