---
license: mit
pipeline_tag: image-to-image
---

# ReChannel: Pixel-Space Dense Prediction with Text-to-Image Models

This repository contains the weights for **ReChannel** presented in [From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models](https://huggingface.co/papers/2607.06553).

ReChannel reads dense prediction targets (such as depth, surface normals, matting, and referring segmentation) out of a FLUX-Klein text-to-image DiT with lightweight per-task LoRA adapters on the (otherwise frozen) backbone. 

* **Code:** [GitHub Repository](https://github.com/xmz111/ReChannel)

## Quick Start (Inference)

To run the model locally, please clone the GitHub repository and install the dependencies:

```bash
git clone https://github.com/xmz111/ReChannel.git
cd ReChannel
pip install -r requirements.txt
```

You can then run dense prediction tasks on a single image:

```bash
python infer.py --image assets/demo_input.jpg \
    --tasks depth,normal,matting,refseg \
    --phrase "the right couch" \
    --out out.png
```

- `--tasks`: any subset of `depth, normal, matting, refseg`.
- `--phrase`: the referring expression used by `refseg` (text-conditioned).

## How it works

```
RGB --VAE encoder--> latent tokens --DiT (frozen θ + task LoRA Δt, σ=0)--> token field Z_t
    --  Ŷ = reshape( W_t · z_ij + b_t ) ∈ R^{p×p×K}  --tile over the plane-->  dense field
```

The backbone is frozen; only a lightweight per-task LoRA adapter and the token-local linear head are trained. The head has no spatial mixing — all spatial structure comes from the adapted token field, not the head.

## Citation

```bibtex
@article{wang2026rechannel,
  title={From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models},
  author={Wang, Zanyi and Lin, Xin and Li, Haodong and Jiang, Dengyang and Li, Yijiang},
  journal={arXiv preprint arXiv:2607.06553},
  year={2026}
}
```