SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
SoL-Refiner converts low-resolution outputs into 4K video in one denoising step, beating multi-step refiners.
SoL-Refiner is a one-step video refiner that turns low-resolution generator outputs into 4K video with a single denoising step. Its recipe combines high-resolution continual training, reinforcement-learning post-training, and one-step distillation, evaluated on Refiner-Bench. At about 2K it leads external refiners on VBench and UniPercept averages, and at 3840x2176 it improves both metrics over the three-step LTX-2.3 Refiner. The full acceleration stack reports an 8.91x refinement-latency speedup versus that baseline in the 2K setting.
- One denoising step upscales low-resolution video to 4K.
- Training combines continual pretraining, RL post-training, and one-step distillation.
- At 2K it leads external refiners on VBench and UniPercept.
- At 3840x2176 it beats the three-step LTX-2.3 Refiner.
- Full stack yields an 8.91x refinement-latency speedup at 2K.
- AI models
- SoL-RefinerLTX-2.3
From the original article
1 min read at huggingface.co
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.
By Haozhe Liu, Tian Ye, Shuchen Xue · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.