📋 Task Overview
This demo supports six tasks. Select one to get started:
| Task | Description |
|---|---|
| Prompt Tuning – Image Editing | Generate an edit using Flux (Kontext) from a source image and instruction, then evaluate it. Use the refinement to tune your prompt. |
| Prompt Tuning – T2I Generation | Upload a generated image and a text-to-image prompt, then evaluate it. Use the refinement to iteratively improve your prompt. |
| Pointwise – Image Editing | Rate a single edited image against its source image and the editing instruction. Produces per-aspect scores and a refined request. |
| Pointwise – T2I Generation | Rate a single generated image against a text-to-image prompt. Produces per-aspect scores and a refined prompt. |
| Pairwise – Image Editing | Compare two edited images (A vs B) given a source image and editing instruction. Determines which edit is better per aspect. |
| Pairwise – T2I Generation | Compare two generated images (A vs B) given a text-to-image prompt. Determines which generation is better per aspect. |
Try the examples on the right - they're basically begging to be clicked! 🎯
| Task Type | Editing Instruction | Source Image | Generated Image | Image B |
|---|
Terms of use
By using this service, users are required to agree to the following terms: The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research.
We deploy our model backend with SGLang. However, there could be congestion during the serving process, leading to delayed responses. If you encounter any issues with the webpage, kindly refresh it.
License
The service is a research preview and is subject to the License of Qwen2, the License of LLaVA-NEXT, and the Terms of Use governing the data generated by OpenAI. Users are required to strictly adhere to the terms outlined in these licenses. Please contact us if you identify any potential violations.
Citation
@article{RationalRewards,
title={RationalRewards: Empowering Image Synthesis with Rationalized Reward Models},
author = {Wang, Haozhe and Qu, Chao and Huang, Zuming and Chu, Wei and Lin,Fangzhen and Chen, Wenhu},
journal={arXiv preprint arXiv:2504.08837},
year={2025}
}