✦
Summary
SoL-Refiner enhances lower-resolution generated videos with a single high-resolution refinement step. High-resolution continual training, frame-based reinforcement learning, and distribution-matching distillation produce a one-step refiner for multiple base generators.
About SoL-Refiner
Video developers can separate economical base generation from detail recovery instead of paying for repeated full-resolution denoising. Core capabilities include Low-resolution video input, Up to 4K refinement, One target-resolution denoising step, and Cross-generator support.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Frame-based RL post-training, One-step distillation, Bidirectional inference, and Streaming inference.