HKUST-C4G / diffusion-rm

What is diffusion-rm?

diffusion-rm is a Python toolkit for building diffusion-native reward models that score image latents directly during generation, skipping costly VLM evaluations for cheaper alignment. Plug it into Stable Diffusion 3.5 pipelines to train on preference pairs or run inference on generated latents or local images via simple API calls like scorer.reward(text_conds, latents, u=0.4). It delivers normalized scores out-of-the-box, with pretrained models on Hugging Face.

Why is it gaining traction?

It matches VLM reward accuracy at lower compute—key for scaling diffusion alignment beyond 20 github prototypes—while offering plug-and-play inference that integrates with diffusers pipelines. Clear examples for scoring on-the-fly latents or disk images lower the barrier versus custom reward hacks. Research from HKUST shows competitive Thurstone loss performance, hooking devs tired of VLM latency.

Who should use this?

Diffusion fine-tuners aligning SD3 outputs to human prefs without VLM overhead. RLHF teams prototyping reward signals for image gen pipelines. Researchers beyond github nanogpt baselines needing diffusion RM for custom backbones like Flux.

Verdict

Grab it for inference today—pretrained model works seamlessly—but skip full training until dataset and eval code drop per the roadmap. 41 stars and 1.0% credibility score signal early-stage research code; solid docs make it usable now for diffusion RM experiments.

(198 words)