Quick answer (checked 12 October 2026): SoL-Refiner ("Speed-of-Light One-Step Refinement for High-Resolution Video") is a research model from NVIDIA Research (Efficient AI Team and Singapore Lab), published on arXiv on 29 September 2026. Instead of generating every frame at full resolution, you generate a cheap low-resolution draft with any video model, then SoL-Refiner refines it to high resolution, up to 4K, in a single denoising step. Code is in NVIDIA's Sana repository and weights are on Hugging Face.
High-resolution AI video is slow because cost grows quickly with the number of pixels and frames. A common workaround is "generate small, then upscale", but most refiners need several sampling steps, which creates a second bottleneck. SoL-Refiner's idea is to do that refinement in one step. Everything below comes from NVIDIA's project page, paper and code repository (linked at the end). I haven't run it myself; the reported numbers come from NVIDIA's own measurements on data-center GPUs.
How it works
The pipeline has two stages:
- Draft: a base video generator produces a lower-resolution video. NVIDIA demonstrates SANA-Video, WAN, Cosmos-Nano and MiniMax H3.
- Refine once: SoL-Refiner takes that draft and restores detail at the target resolution in a single denoising step.
The refiner is trained in three stages, according to the paper:
- High-resolution continual training on aligned pairs of low- and high-quality videos.
- Frame-based reinforcement-learning post-training with visual reward feedback to improve perceived quality.
- One-step distribution-matching distillation, which compresses a multi-step "teacher" refiner into a one-step "student".
The original paper builds on the LTX-2.3 Refiner as its backbone. NVIDIA later released a separate LTX-2.5-based variant for MiniMax H3, which the repository notes is outside the paper's experiments. NVIDIA also introduced Refiner-Bench, a benchmark of outputs from different video generators, to compare refiners fairly.
Reported results
All of these are NVIDIA's own measurements, with the hardware they used:
| Setup (NVIDIA's protocol) | Without SoL-Refiner | With draft + one-step refine | Speedup |
|---|---|---|---|
| MiniMax H3, 5-second 1344×768 video at 24 fps, GB200 GPUs | 152.3 s at full resolution | 5.64 s total (4.06 s draft + 1.56 s refine) | 27.03× |
| WAN, 81 frames at 24 fps, 1× H100 | 272.69 s direct at 1280×720 | 78.71 s at 1280×704 | 3.46× |
| Cosmos-Nano (35-step), 189 frames at 24 fps, 1× H100 | n/a | n/a | 2.81× |
| SANA-Video, 81 frames at 16 fps, 1× H100 | 20.04 s direct at 1248×704 | 12.54 s (832×480 + refine) | 1.60× |
On quality, the paper reports that at about 2K the one-step SoL-Refiner beats all the external refiners it tested on VBench and UniPercept averages, and that at 3840×2176 it improves both metrics over the three-step LTX-2.3 Refiner. With its full acceleration stack, NVIDIA reports an 8.91× speedup in refinement latency over the same baseline in its 2K setting.
Two cautions. These timings are on H100 and GB200 data-center GPUs, not consumer cards. And the repository notes that the project-page figures are research results, separate from benchmarks of the released Diffusers implementation.
Why it matters
- It works across generators. The refiner and the base model are separate choices, so a studio could keep its preferred generator and add a refinement stage.
- Cheaper high-resolution output. Generating at lower resolution and refining once can cut generation time substantially, according to the measurements above.
- Open research artefacts. The code and weights are public, so others can reproduce or build on the results.
Limitations to keep in mind
- Refining is not the same as the original. Any refiner adds detail that wasn't in the low-resolution draft. For creative work that's usually the point; for anything where faithfulness matters, such as product footage or evidence, compare outputs carefully.
- Heavy hardware. The published results use data-center GPUs; NVIDIA hasn't published consumer-GPU requirements.
- Research release. This is a research project, not a packaged consumer app.
Where to get it
- Code: the
sol-refinerfolder in NVIDIA's Sana repository, with separate instructions for the LTX-2.3 paper version and the MiniMax-H3 extension. - Weights on Hugging Face: SoL-Refiner LTX-2.3 One-Step, LTX-2.3 Multi-Step and LTX-2.5 for MiniMax-H3.
Check each model card for its license and requirements before using it in a product.
Frequently asked questions
What does SoL stand for?
"Speed-of-Light", from the paper's title: SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video.
Does SoL-Refiner generate video by itself?
No. It refines a video made by another generator, such as WAN, Cosmos-Nano, SANA-Video or MiniMax H3.
Can I run it on a gaming GPU?
NVIDIA's published results use H100 and GB200 GPUs and it hasn't published consumer-GPU requirements, so check the repository instructions before trying.
Sources
- NVIDIA Research: SoL-Refiner project page
- arXiv: SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video (2609.37969)
- GitHub: NVlabs/Sana, sol-refiner
Related on VNHAX: ByteDance DMAD: 4-step video generation · ByteDance PDMD explained · NVIDIA PixelUMM explained
