Andrei Bursuc @CVPR#1697
Excited to share that MIRO is accepted to ICML 2026 @icmlconf ! 🎉
We introduce multi-reward conditioned training for text-to-image. By training on continuous reward scores, we can simply condition on HIGH REWARDS at inference to guarantee top-tier, aligned outputs.
We’ve open-sourced everything, including all individual reward-ablated model variants!
🌐 Site: https://nicolas-dufour.github.io/miro/ 📄 Paper: https://arxiv.org/abs/2510.25897 🛠️ Git: https://github.com/nicolas-dufour/miro 🤗 HF: https://huggingface.co/nicolas-dufour/miro 🎨 Demo: https://huggingface.co/spaces/nicolas-dufour/miro
First, a reminder of MIRO's principal results and efficiency:
⚡ Up to 19× faster training convergence than standard baselines. 📉 34× fewer parameters & 370× cheaper inference compute than models like FLUX, while maintaining competitive visual quality.
Are all 7 rewards actually useful? Yes!
Our new "leave-one-out" ablation shows that removing even a single reward drops overall performance. Even though these rewards are quite entangled, each one still injects unique, useful bits of info.
By conditioning on a vector of 7 rewards simultaneously, MIRO naturally balances conflicting objectives and avoids reward hacking.
This yields a major jump in text composition, achieving SOTA scores on GenEval, PickAScore, and HPSv2.
We can control at test time the reward mix
We also expand MIRO beyond training from scratch, it works also as a post-training framework!
Applying multi-reward conditioning during fine-tuning on an existing base model yields the a robust, controllable alignment at inference.
We’ve added a firm mathematical foundation.
Our new theorem proves that conditioning on the joint reward distribution guarantees the model steers toward high-reward regions while preserving sample diversity and avoiding reward hacking.
Work done with @lucasdegeorge , @sohonjitghosh, @VickyKalogeiton and @david_picard
Andrei Bursuc @CVPR#1697Excited to share that MIRO is accepted to ICML 2026 @icmlconf ! 🎉
We introduce multi-reward conditioned training for text-to-image. By training on continuous reward scores, we can simply condition on HIGH REWARDS at inference to guarantee top-tier, aligned outputs.