---
license: mit
library_name: miro-t2i
tags:
  - text-to-image
  - diffusion
  - flow-matching
  - miro
  - reward-conditioning
  - ablations
pipeline_tag: text-to-image
---

# MIRO — ablations and single-reward specialists

This repository hosts the **15 ablation / baseline checkpoints** that accompany the main MIRO release at [`nicolas-dufour/miro`](https://huggingface.co/nicolas-dufour/miro).

> Dufour, Degeorge, Ghosh, Kalogeiton, Picard. _MIRO: MultI-Reward cOnditioned pretraining improves T2I quality and efficiency_. **ICML 2026**.
>
> 📄 [Paper](https://arxiv.org/abs/2510.25897) · 🌐 [Project page](https://nicolas-dufour.github.io/miro/) · 💻 [Code](https://github.com/nicolas-dufour/miro) · 🐍 `pip install miro-t2i`

<table style="width:100%;border-collapse:separate;border-spacing:4px">
<tr><td><img src="https://huggingface.co/nicolas-dufour/miro/resolve/main/teaser.jpg" alt="MIRO samples"></td></tr>
</table>

## Layout

Every variant lives in its own subfolder and is loaded via the `variant=` argument:

```python
from miro import MiroPipeline
import torch

pipe = MiroPipeline.from_pretrained(
    "nicolas-dufour/miro-ablations",
    variant="miro-no-clip",            # ← the subfolder name
).to("cuda", torch.float16)
```

Each `MiroPipeline` instance exposes `pipe.coherence_keys`, which lists the reward axes the loaded checkpoint was trained on. `reward_targets={...}` will raise `ValueError` if you pass a key that's not in this list.

## Variants

### Reward ablations (8) — full MIRO recipe minus one signal

Same architecture and training data as the main MIRO, with one reward signal turned off so you can isolate its contribution.

| Subfolder | What's ablated | `coherence_keys` size |
|---|---|:-:|
| `miro-no-synthetic-captions` | Trained on original captions only (no synthetic-caption augmentation) | 7 |
| `miro-no-aesthetic` | LAION aesthetic-quality reward | 6 |
| `miro-no-clip` | CLIP text-image alignment | 6 |
| `miro-no-hpsv2` | HPSv2 human preference | 6 |
| `miro-no-image-reward` | ImageReward | 6 |
| `miro-no-pickscore` | PickScore human preference | 6 |
| `miro-no-sciscore` | SciScore | 6 |
| `miro-no-vqa` | VQAScore | 6 |

### Single-reward specialists (7) — paper baselines

Each is trained on **only one** reward signal — the controls the paper compares MIRO against. `pipe.coherence_keys` is a 1-tuple for these.

| Subfolder | The one reward it knows about |
|---|---|
| `miro-only-aesthetic` | `aesthetic_score` |
| `miro-only-clip` | `clip_score` |
| `miro-only-hpsv2` | `hpsv2_score` |
| `miro-only-image-reward` | `image_reward_score` |
| `miro-only-pickscore` | `pick_a_score_score` |
| `miro-only-sciscore` | `sciscore_score` |
| `miro-only-vqa` | `vqa_score` |

## What's in each subfolder

```
miro-<variant>/
├── model.safetensors      # fp32 EMA weights (~1.4 GB) — ready for finetuning
├── config.json            # network kwargs + sampler defaults
├── uncond_embedding.npy   # precomputed FLAN-T5-XL unconditional embedding
├── teaser.jpg             # shared masonry gallery
└── README.md              # per-variant model card
```

## Citation

```bibtex
@inproceedings{dufour2026miro,
  title     = {{MIRO}: {M}ult{I}-{R}eward c{O}nditioned pretraining improves {T2I} quality and efficiency},
  author    = {Dufour, Nicolas and Degeorge, Lucas and Ghosh, Arijit and Kalogeiton, Vicky and Picard, David},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```

## License

MIT — see [LICENSE](https://github.com/nicolas-dufour/miro/blob/main/LICENSE).