|
|
--- |
|
|
library_name: diffusers |
|
|
license: mit |
|
|
pipeline_tag: text-to-image |
|
|
base_model: |
|
|
- black-forest-labs/FLUX.2-klein-base-9B |
|
|
--- |
|
|
|
|
|
# Model Summary |
|
|
This model is GRPO trained using [UnifiedReward-Flex](https://huggingface.co/collections/CodeGoat24/unifiedreward-flex) as reward on the training dataset of [UniGenBench](https://github.com/CodeGoat24/UniGenBench). |
|
|
|
|
|
🚀 The inference code is available at [Github](https://github.com/CodeGoat24/Pref-GRPO/blob/main/inference/flux2_klein_dist_infer.sh). |
|
|
|
|
|
|
|
|
For further details, please refer to the following resources: |
|
|
- 📰 Paper: https://arxiv.org/abs/2602.02380 |
|
|
- 🪐 Project Page: https://codegoat24.github.io/UnifiedReward/flex |
|
|
- 🤗 Model Collections: https://huggingface.co/collections/CodeGoat24/unifiedreward-flex |
|
|
- 🤗 Dataset: https://huggingface.co/datasets/CodeGoat24/UnifiedReward-Flex-SFT-90K |
|
|
- 👋 Point of Contact: [Yibin Wang](https://codegoat24.github.io) |
|
|
|
|
|
# Qualitative Results |
|
|
 |
|
|
|
|
|
|
|
|
|
|
|
# Quantitative Results |
|
|
|
|
|
## UniGenBench |
|
|
| Model | Overall | Style | World Knowledge | Attribute | Action | Relationship | Compound | Grammar | Logical Reasoning | Entity Layout | Text Generation | |
|
|
|:------|:------:|:-----:|:---------------:|:---------:|:------:|:------------:|:--------:|:-------:|:----------------:|:------------:|:--------------:| |
|
|
| FLUX2.Klein-base-9B | 78.93% | 97.50% | 91.61% | 83.65% | 77.00% | 86.42% | 78.61% | 76.87% | 53.41% | 88.43% | 55.75% | |
|
|
| **Ours** | **81.54%** | **97.60%** | **91.93%** | **85.47%** | **78.42%** | **86.42%** | **81.96%** | **76.97%** | **58.64%** | **88.43%** | **69.54%** |
|
|
|
|
|
|
|
|
## T2I-CompBench |
|
|
| Model | Overall | Color | Shape | Texture | 2D-Spatial | 3D-Spatial | Numeracy | Non-Spatial | Complex | |
|
|
|--------------------------|--------:|-------:|-------:|--------:|-----------:|-----------:|---------:|------------:|--------:| |
|
|
| FLUX2.Klein-base-9B | 53.72% | 85.90% | 60.81% | 72.24% | 41.46% | 36.87% | 64.36% | 31.11% | 37.04% | |
|
|
| **Ours**| **58.75%** | **85.93%** | **63.36%** | **74.69%** | **46.77%** | **43.18%** | **70.60%** | **30.73%** | **54.73%** | |
|
|
|
|
|
|
|
|
## GenEval |
|
|
| Model | Overall | Single Object | Two Object | Counting | Colors | Position | Color Attr | |
|
|
|--------------|---------:|--------------:|-----------:|---------:|-------:|---------:|-----------:| |
|
|
| FLUX2.Klein-base-9B | 78.99% | 99.69% | 92.93% | 77.50% | 92.55% | 66.75% | 44.50% | |
|
|
| **Ours** | **81.55%** | **99.69%** | **93.94%** | **84.69%** | **93.22%** | **70.75%** | **47.00%** | |
|
|
|
|
|
|
|
|
|
|
|
## Citation |
|
|
|
|
|
```bibtex |
|
|
@article{unifiedreward-flex, |
|
|
title={Unified Personalized Reward Model for Vision Generation}, |
|
|
author={Wang, Yibin and Zang, Yuhang and Han, Feng and Bu, Jiazi and Zhou, Yujie and Jin, Cheng and Wang, Jiaqi}, |
|
|
journal={arXiv preprint arXiv:2602.02380}, |
|
|
year={2026} |
|
|
} |
|
|
``` |
