NVIDIA SoL-Refiner: 4K video changes method
6 octobre 2026 · Kris Wilhem
NVIDIA Research has just presented SoL-Refiner, an approach designed to speed up high-resolution video refinement.
The idea is simple on paper: instead of producing each frame directly at high definition, the pipeline first generates a lighter version more quickly, then applies a single refinement step at the final resolution.
On MiniMax H3, NVIDIA announces in its own measurements a drop from 152.3 seconds to 5.64 seconds, about 27× faster. Those are NVIDIA's figures, not yet ours.
What really interests us: real integration
At Axiris, we have already installed SoL-Refiner in our test environment and launched the first comparative renders against our current 4K upscale chain based on SeedVR2.
And that is where the subject becomes really interesting.
To run SoL-Refiner, we could not keep ComfyUI 0.38, the latest version used in our main environment. The project's dependencies around Triton forced us to use an earlier version of ComfyUI in an isolated lab.
This kind of constraint is rarely visible in an official announcement, but it becomes central as soon as a new model has to actually enter an existing production chain.
Our first tests: a reconstruction that is sometimes smoother
On our first renders, we immediately noticed a particular behaviour: when SoL has to invent material in zones where the source does not carry enough information, the result can tend to be a little smoother.
We had spotted this impression visually before even going back to the documentation. The node's author does warn that refinement can alter certain faces, patterns and details.
That is precisely the kind of behaviour we want to measure.
An upscale is not judged only by the number of extra pixels. You have to look at how the material is rebuilt, whether textures stay credible, whether faces stay stable and whether the result really respects the original shot.
SoL against SeedVR2
Our benchmark is now under way on our own sequences.
The question is simple:
Does SoL-Refiner really improve our 4K output compared with our current SeedVR2 pipeline?
We will compare the quality of detail, the stability of textures and faces, fidelity to the original film, compute times and the real value of the final result.
At Axiris, this is exactly what we expect from useful technology watch: install, confront, measure, then integrate a new technology into a real production chain, or not.
What this changes for studios
Above all, SoL-Refiner shows that the future of AI video will not run only through ever-heavier models.
It will also run through intelligent refinement chains, able to produce faster and then improve the final definition.
But between a demonstration and a production tool, one essential question remains: is it stable, faithful and integrable enough to replace what already works?
That is what we are in the process of verifying.