MIRO — Multi-reward conditioned text-to-image

Interactive demo for MIRO (ICML 2026): drag the seven reward sliders to steer generation toward each axis (CLIP alignment, aesthetic quality, ImageReward, PickScore, HPSv2, VQAScore, SciScore).

• 📄 Paper
• 🌐 Project page
• 💻 Code
• 🤗 Models
• 🐍 PyPI — pip install miro-t2i

The demo runs on ZeroGPU H200: the GPU is allocated on-demand per inference call. First request takes ~10s of cold-start; subsequent ones are typically <5s for 4 samples at 50 steps.