Yijun Liu commited on
Commit ·
3eb7599
1
Parent(s): ac26578
Add resource badges and standardize the paper citation
Browse files- README.md +20 -5
- citation.bib +8 -0
README.md
CHANGED
|
@@ -3,20 +3,30 @@ license: apache-2.0
|
|
| 3 |
base_model:
|
| 4 |
- Qwen/Qwen3-VL-8B-Instruct
|
| 5 |
pipeline_tag: image-text-to-text
|
|
|
|
|
|
|
| 6 |
tags:
|
| 7 |
- reward-model
|
| 8 |
- text-to-image
|
| 9 |
- human-preference
|
| 10 |
- rlhf
|
|
|
|
| 11 |
---
|
| 12 |
|
| 13 |
# HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities
|
| 14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
HPSv3++ is a **capability-aware and RL-iteration-aware** text-to-image (T2I) reward model, built on the `Qwen/Qwen3-VL-8B-Instruct` backbone with a Capability Encoder, a FiLM conditioning head, and a three-layer RankNet reward head.
|
| 16 |
|
| 17 |
A Capability Encoder implicitly infers the generative ability of the model that produced an image, while the RL iteration step is supplied as an explicit condition; the two are jointly modulated through FiLM so that a single reward model produces calibrated preference scores across generators of differing capability and different stages of RL optimization.
|
| 18 |
|
| 19 |
-
|
| 20 |
|
| 21 |
## Files
|
| 22 |
|
|
@@ -35,10 +45,15 @@ The training/evaluation dataset, HPDv3++, is released separately: [Junjun2333/HP
|
|
| 35 |
|
| 36 |
## Citation
|
| 37 |
|
|
|
|
|
|
|
| 38 |
```bibtex
|
| 39 |
-
@
|
| 40 |
-
title
|
| 41 |
-
author
|
| 42 |
-
|
|
|
|
|
|
|
|
|
|
| 43 |
}
|
| 44 |
```
|
|
|
|
| 3 |
base_model:
|
| 4 |
- Qwen/Qwen3-VL-8B-Instruct
|
| 5 |
pipeline_tag: image-text-to-text
|
| 6 |
+
datasets:
|
| 7 |
+
- Junjun2333/HPDv3-PlusPlus
|
| 8 |
tags:
|
| 9 |
- reward-model
|
| 10 |
- text-to-image
|
| 11 |
- human-preference
|
| 12 |
- rlhf
|
| 13 |
+
- arxiv:2606.14657
|
| 14 |
---
|
| 15 |
|
| 16 |
# HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities
|
| 17 |
|
| 18 |
+
<p align="center">
|
| 19 |
+
<a href="https://arxiv.org/abs/2606.14657"><img src="https://img.shields.io/badge/arXiv-2606.14657-b31b1b.svg" alt="Paper: arXiv 2606.14657"></a>
|
| 20 |
+
<a href="https://github.com/PlantPotatoOnMoon/HPSv3-PlusPlus"><img src="https://img.shields.io/badge/GitHub-Code-181717?logo=github" alt="GitHub: HPSv3++ code"></a>
|
| 21 |
+
<a href="https://huggingface.co/datasets/Junjun2333/HPDv3-PlusPlus"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-HPDv3%2B%2B-yellow" alt="Dataset: HPDv3++"></a>
|
| 22 |
+
<a href="https://huggingface.co/Junjun2333/HPSv3-PlusPlus/resolve/main/citation.bib?download=true"><img src="https://img.shields.io/badge/Citation-BibTeX-2563eb" alt="Citation / BibTeX"></a>
|
| 23 |
+
</p>
|
| 24 |
+
|
| 25 |
HPSv3++ is a **capability-aware and RL-iteration-aware** text-to-image (T2I) reward model, built on the `Qwen/Qwen3-VL-8B-Instruct` backbone with a Capability Encoder, a FiLM conditioning head, and a three-layer RankNet reward head.
|
| 26 |
|
| 27 |
A Capability Encoder implicitly infers the generative ability of the model that produced an image, while the RL iteration step is supplied as an explicit condition; the two are jointly modulated through FiLM so that a single reward model produces calibrated preference scores across generators of differing capability and different stages of RL optimization.
|
| 28 |
|
| 29 |
+
**[HPDv3++](https://huggingface.co/datasets/Junjun2333/HPDv3-PlusPlus)** is an independently usable human preference training dataset and benchmark with approximately **212K preference pairs** covering text fidelity and aesthetic quality. See the [dataset guide](https://github.com/PlantPotatoOnMoon/HPSv3-PlusPlus#dataset-hpdv3).
|
| 30 |
|
| 31 |
## Files
|
| 32 |
|
|
|
|
| 45 |
|
| 46 |
## Citation
|
| 47 |
|
| 48 |
+
Please cite the paper when using the HPSv3++ model or HPDv3++ dataset. [Download BibTeX](https://huggingface.co/Junjun2333/HPSv3-PlusPlus/resolve/main/citation.bib?download=true).
|
| 49 |
+
|
| 50 |
```bibtex
|
| 51 |
+
@article{liu2026hpsv3plusplus,
|
| 52 |
+
title = {HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities},
|
| 53 |
+
author = {Liu, Yijun and Huang, Jie and Xue, Zeyue and Li, Yuming and He, Ruizhe and Li, Haoran and Ge, Shijia and Fu, Siming},
|
| 54 |
+
journal = {arXiv preprint arXiv:2606.14657},
|
| 55 |
+
year = {2026},
|
| 56 |
+
doi = {10.48550/arXiv.2606.14657},
|
| 57 |
+
url = {https://arxiv.org/abs/2606.14657}
|
| 58 |
}
|
| 59 |
```
|
citation.bib
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
@article{liu2026hpsv3plusplus,
|
| 2 |
+
title = {HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities},
|
| 3 |
+
author = {Liu, Yijun and Huang, Jie and Xue, Zeyue and Li, Yuming and He, Ruizhe and Li, Haoran and Ge, Shijia and Fu, Siming},
|
| 4 |
+
journal = {arXiv preprint arXiv:2606.14657},
|
| 5 |
+
year = {2026},
|
| 6 |
+
doi = {10.48550/arXiv.2606.14657},
|
| 7 |
+
url = {https://arxiv.org/abs/2606.14657}
|
| 8 |
+
}
|