Tencent Hunyuan's FewStepRL adapts RL fine-tuning to few-step diffusion models
Tencent-Hunyuan · hf · 2026-11-10
Tencent Hunyuan, with Westlake University, introduces FewStepRL, a systematic framework for RL post-training of few-step diffusion models. Existing RL objectives assume instantaneous velocity with many-step samplers, while few-step models model average velocity over intervals, making them a poor fit.
The paper reformulates RL objectives in three phases—x-space transformation, gradient decomposition, and anchor adaptation—yielding modular few-step variants of several methods. Findings: reward-gradient methods excel with differentiable rewards, while DiffusionNFT-style methods achieve high rewards in fewer iterations, cutting post-training cost.
More from Multimodal
- Coder generates a 15-second motion graphics scene with Fable 5.2 using pure code, no video model — aziz4ai · 2026-11-10
- Fable5.2 inside Claude Code recreates a Bora Bora photo almost indistinguishably — AlchainHust · 2026-11-10
- 2D-to-3D character generation demo impresses with faithful style transfer — azed_ai · 2026-11-10
- MOSS-Image-SFT gets control LoRAs: pose, depth, canny and more — pavel_0869874 · 2026-11-10
- 467-run ComfyUI quant benchmark across 10 GPUs: int8 ConvRot most faithful, nvfp4 fastest but drifts — mybrianonacid · 2026-11-10
- AI-Generated Joust Remake Concept Art: Knights on Giant Birds Battle Pterodactyls — PurzBeats · 2026-11-10