---
license: other
library_name: diffusers
pipeline_tag: image-to-image
---

# ICTone-Fill-LoRA

This repository contains the LoRA weights for **ICTone** (ECCV 2026), described in the paper [Towards In-Context Tone Style Transfer with a Large-Scale Triplet Dataset](https://huggingface.co/papers/2604.16114). ICTone performs reference-based tone style transfer by jointly conditioning on the content and reference images with a diffusion transformer (FLUX.1-Fill).

- Project page: [https://dengyuhai.github.io/ICTone_Project/](https://dengyuhai.github.io/ICTone_Project/)
- Code: [https://github.com/dengyuhai/ICTone](https://github.com/dengyuhai/ICTone)
- Dataset: [TST100K](https://huggingface.co/datasets/ToneStyle/TST100K)
- Benchmark: [TST2K](https://huggingface.co/datasets/ToneStyle/TST2K)
- Online demo: [https://huggingface.co/spaces/ToneStyle/ICTone-Fill](https://huggingface.co/spaces/ToneStyle/ICTone-Fill)

## Usage

The following inference example is taken from the GitHub README:

```bash
CUDA_VISIBLE_DEVICES=0 python inference.py \
    --content   assets/example_content.png \
    --reference assets/example_reference.png \
    --flux-path black-forest-labs/FLUX.1-Fill-dev \
    --lora-path ToneStyle/ICTone-Fill-LoRA \
    --output-file ./output/example_output.png \
    --num-inference-steps 4 \
    --guidance-scale 50.0 \
    --seed 666
```

For full training and evaluation instructions, please refer to the [GitHub repository](https://github.com/dengyuhai/ICTone).

## License

Non-commercial research use only. See the repository and paper for details.