Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
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 |
+
The training/evaluation dataset, HPDv3++, is released separately: [Junjun2333/HPDv3-PlusPlus](https://huggingface.co/datasets/Junjun2333/HPDv3-PlusPlus).
|
| 20 |
+
|
| 21 |
+
## Files
|
| 22 |
+
|
| 23 |
+
| File | Description |
|
| 24 |
+
|---|---|
|
| 25 |
+
| `hpsv3++.pth` | Final HPSv3++ reward-model weights (17.6 GB) |
|
| 26 |
+
| `config.json` | Model configuration |
|
| 27 |
+
|
| 28 |
+
## Conditioning at inference
|
| 29 |
+
|
| 30 |
+
- **Model capability** is inferred implicitly from the image; you do not pass it in.
|
| 31 |
+
- **RL iteration** is passed explicitly as a normalized scalar in `[0, 1]`.
|
| 32 |
+
- General preference scoring / ranking: use `0.0` (pre-RL setting).
|
| 33 |
+
- As the reward inside T2I RL fine-tuning: ramp the iteration condition linearly from `0.3` to `1.0` over training (the setting used in the paper).
|
| 34 |
+
- Use the mean (`mu`) output as the scalar reward.
|
| 35 |
+
|
| 36 |
+
## Citation
|
| 37 |
+
|
| 38 |
+
```bibtex
|
| 39 |
+
@misc{hpsv3pp,
|
| 40 |
+
title = {HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities},
|
| 41 |
+
author = {HPSv3++ Team},
|
| 42 |
+
year = {2026}
|
| 43 |
+
}
|
| 44 |
+
```
|