VL-DNP

1 Set-up

# create / activate your venv or conda env first
pip install -r requirements.txt

#  (Optional) install a GPU wheel if you have CUDA 12.1
pip install --upgrade torch --index-url https://download.pytorch.org/whl/cu121

The pinned versions avoid the numpy 2.0 / transformers padding bug that throws “expected np.ndarray (got numpy.ndarray)”.

---

2 Running Diffusion Model with VLM

VL-DNP is based on Qwen 2.5-VL-7B-Instruct and Stable Diffusion v1.4

# VLM evaluating during Diffusion sampling.
python code/dpm_with_VLM.py \
	--path results \
	--vlm_step 5 6 7 9 12 16 21 27 34 42 \
	--obj ring-a-bell-16 \
	--neg_guidance 15

• --path directory to save generated image.
• --vlm_step steps that VLM generates negative prompt.
• --obj evaluating prompt set. coco is for normal prompts. ring-a-bell is for adversarial prompts.
• --neg_guidance negative guidance scale to be used

---

3 Running Diffusion Model with Static Negative Prompting

# negative prompting evaluation using various negative guidance sclae
python code/negative_prompt.py \
	--path results_neg_prompt \
	--neg_guidance 15 \
	--obj ring-a-bell-16

• --path directory to save generated image.
• --obj evaluating prompt set. coco is for normal prompts. ring-a-bell is for adversarial prompts.
• --neg_guidance negative guidance scale to be used

---

4 Evaluation with Nudenet Classifier (from SAFREE)

You can download Classifier model at Nudenet Classifier

Download and place at 'classifier/'

Evaluation can be done by

# Nudity evaluation using Nudenet Classifier
python code/nudity_eval.py \
	--dir ./results/dir

• --dir directory of images to be evaluated. After evaluation, it will output Attack Success Rate and Toxic Rate.

5 Adversarial Prompt Set List

Adversarial Prompt Sets are from

1. UnlearnAtk
2. P4D
3. MMA-Diff
4. Ring-A-Bell

Download and place at 'prompt_set/'

6 Citation

If you find this code useful for your research, please cite as follows:

@misc{chang2025dynamicvlmguidednegativeprompting,
      title={Dynamic VLM-Guided Negative Prompting for Diffusion Models}, 
      author={Hoyeon Chang and Seungjin Kim and Yoonseok Choi},
      year={2025},
      eprint={2510.26052},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.26052}, 
}

Paper Link