SoL-Refiner is a one-step video refiner that upsamples low-resolution videos to high resolution (up to 4K) using a single denoising step. It is trained in three stages: continual training on high-resolution videos, post-training with frame-based rewards, and one-step distillation. It works with multiple base video generators without retraining them. On the newly introduced Refiner-Bench, SoL-Refiner outperforms all external refiners on VBench and UniPercept averages at 2K resolution. At 3840x2176, it improves over the three-step LTX-2.3 refiner on both metrics. The complete acceleration stack achieves an 8.91x speedup in refinement latency compared to the three-step baseline. This makes high-resolution video generation more efficient by moving most computation to a lower-resolution base generator and reserving a single refinement step.
One-step video refiner achieves 8.91x speedup for 4K upscaling
SoL-Refiner uses a three-stage training recipe to refine low-resolution video to 2K/4K in a single denoising step, outperforming multi-step refiners on perceptual quality metrics.
Big Tech
Haozhe Liu · Tian Ye · Shuchen Xue · Yitong Li · Junsong Chen · Haopeng Li · +6 more
NVIDIA
Research Digest··2 min read
In one line
SoL-Refiner transforms low-resolution video outputs into 4K videos with a single denoising step, outperforming multi-step refiners.
Before you read this
Three ideas this paper builds ondistribution matching distillationA training method that aligns student and teacher output distributions to compress many-step generative models into efficient few-step versions.29 papers →flow matchingA generative training method that learns an ordinary differential equation mapping noise to data by matching velocity fields along probability paths.175 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.301 papers →
Read it as
SoL-Refiner is a one-step video refiner that takes low-resolution outputs from various generators and produces high-resolution videos (up to 4K) with a single denoising step.
The authors start from a pretrained LTX-2.3 checkpoint and apply a three-stage recipe: high-resolution continual training on paired videos, RL post-training with frame-based reward models, and a final one-step distillation using a projected DiT discriminator.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.