FLUX.1 SRPO [dev] is a 12 billion parameter flow transformer that generates high-quality images from text with incredible aesthetics. It is a FLUX model fine-tuned with Tencent's SRPO technique.
The paper “Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference” shows a small boost in image quality.
Base FLUX.1 Dev scored 70.8% for excellent and 89.27% for excellent plus good on how well images match prompt. The SRPO-tuned version bumps that up to 73.2% and 90.33%.
The main change is in realism. FLUX got 8.2% for excellent and 64.33% for excellent plus good. SRPO jumps to 38.9% and 80.86%.
Self-hosting SRPO = non-commercial only (unless you obtain a special license from Tencent).
However if you're paying for use through API means commercial use can be allowed, because you’re licensing the service from the provider, not the raw model from Tencent. The API provider usually has a commercial agreement in place with Tencent.
If you'd like to access this model, you can explore the following possibilities:
Spaces