Research map / Image generation / Text-to-image diffusion
MIRO: MultI-Reward cOnditioned pretraining improves T2I quality and efficiency
Nicolas Dufour, Lucas Degeorge, Arijit Ghosh, Vicky Kalogeiton, David Picard
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition on seven rewards during pretraining instead of no reward conditioning | 0.36B flow matching T2I model · GenEval (real captions) | overall: 52 -> 57 | Improved |
| condition on multiple rewards instead of aesthetic reward only (ablation) | MIRO · GenEval (real captions) | overall: 33 -> 57 | Improved |
| condition on multiple rewards instead of one alignment-related reward (CLIP, VQA) (ablation) | MIRO · GenEval (real captions) | overall: 57 -> 57 | No clear effect |
| condition on multiple rewards instead of single reward (ablation) | MIRO · CC12M+LA6 reward metrics | AestheticScore, PickScore, ImageReward, HPSv2 | Improved |
| add reward conditioning to pretraining | 0.36B flow matching T2I model · CC12M+LA6 training convergence | steps to reach baseline performance: 19x faster AestheticScore, 3.3x ImageReward | Improved |
| add synthetic captions to MIRO | MIRO · GenEval | overall: 57 -> 68 | Improved |
| set aesthetic positive target to 0.625 at inference (ablation) | MIRO (synthetic captions) · GenEval | overall: 68 -> 75 | Improved |
| set aesthetic positive target to 0.625 at inference (ablation) | MIRO (synthetic captions) · PartiPrompts | Aesthetic score: 6.28 -> 5.24 | Hurt |
| add best-of-N test-time scaling (128 samples) with ImageReward selection | MIRO (synthetic captions) · GenEval | overall: 68 -> 75 | Improved |
| select samples by aesthetic score with 128-sample scaling | MIRO (synthetic captions) · GenEval | overall: 68 -> 63 | Hurt |
| use MIRO instead of baseline under best-of-N sampling | MIRO · test-time scaling on ImageReward and PickScore | samples needed to match baseline at 128: 8 samples (ImageReward), 4 samples (PickScore) | Improved |
| condition on multiple rewards instead of aesthetic reward only (ablation) | MIRO · PartiPrompts | ImageReward: 0.00 -> 1.06 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Sana-1.6B — SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.87)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Policy Optimized Text-to-Image Pipeline Design (NeurIPS 2025 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.87)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
