---
base_model: Qwen/Qwen2.5-VL-32B-Instruct
library_name: transformers
pipeline_tag: image-text-to-text
license: mit
---

# qwen2vl_32b

This is a PE-finetuned Qwen2.5-VL prompt-enrichment checkpoint for PIPBench, presented in the paper [PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation](https://huggingface.co/papers/2607.06440).

It takes PIPBench reference images plus the original text prompt and outputs an enriched prompt. Generate final images by running Qwen-Image with the enriched prompt.

These weights come from the PIPBench codex/ablations/PE experiments and are not QIP checkpoints.

## Resources

- **Paper:** [PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation](https://huggingface.co/papers/2607.06440)
- **Project Page:** [https://wuyuhang05.github.io/PIPBench/](https://wuyuhang05.github.io/PIPBench/)
- **Repository:** [https://github.com/wuyuhang05/PIPBench](https://github.com/wuyuhang05/PIPBench)

## Usage

To generate enriched prompts using this 32B PE module, clone the [official repository](https://github.com/wuyuhang05/PIPBench) and run:

```bash
bash scripts/enrich_prompts_qwen2vl.sh \
  --model-id AirRain03/qwen2vl_32b \
  --metadata data/pipbench/metadata.json \
  --data-root data/pipbench \
  --out outputs/prompts/qwen2vl_32b.jsonl \
  --device-map auto \
  --dtype bfloat16 \
  --resume
```