Maximilian Augustin*, Yannic Neuhaus*, Matthias Hein
Tübingen AI Center - University of Tübingen
*Equal contribution
This repository contains the official code and resources for our paper:
DASH: Detection and Assessment of Systematic Hallucinations of VLMs
Our object hallucination benchmark DASH-B is already available!
We also provide URLs for ~1 million images resulting from DASH that successfully trigger an object hallucination!
Requirements | Pipeline | Images | Benchmark | arXiv | Citation
conda create --name dash python=3.12
conda activate dash
conda install nvidia/label/cuda-12.1.0::cuda-nvcc
pip install -r requirements_pip.txt
pip install flash-attn
Set the access token in src/dash_utils/vlm_utils/models/hf_access_token.py
conda create -n dash_retrieval python=3.10
conda activate dash_retrieval
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install --no-deps clip-retrieval
pip install -r requirements_retrieval.txt
For retrieval, access to the ReLAION knn-index is also required. We are working on a way to publish it as it is not publicly available at the moment. Stay tuned!
Query Generation
conda activate dash
CUDA_VISIBLE_DEVICES=... python src/1_make_llm_prompts.py source_file=data/spurious_imagenet_class_list.txt result_dir=output/1_prompts/spurious_imagenet
Exploration
conda activate dash_retrieval
ulimit -n 8192
CUDA_VISIBLE_DEVICES=... python src/2_laion_retrieval_prompt_based.py dataset=spurious_imagenet
conda activate dash
CUDA_VISIBLE_DEVICES=... torchrun --nproc-per-node N --standalone src/X_vlm_eval.py source_dir=output/2_prompt_retrieval/spurious_imagenet/llama_only_follow_up model_id=paligemma
Exploitation
conda activate dash_retrieval
CUDA_VISIBLE_DEVICES=... python src/4_exploitation_retrieval.py model_id=paligemma use_agnostic_folder=True
conda activate dash
CUDA_VISIBLE_DEVICES=... torchrun --nproc-per-node N --standalone src/X_vlm_eval.py source_dir=output/4_exploitation_prompt_retrieval/spurious_imagenet/llama_only_follow_up/prompt_4 model_id=paligemma
Optimization
conda activate dash
CUDA_VISIBLE_DEVICES=... torchrun --nproc-per-node N --standalone src/2_sdxl_optim.py gpu=0 diffusion_gpu=1 vae_gpu=1 reg_gpu=0 prompt_type=llama_only_follow_up detection_weight=1.0 model_id=paligemma num_cutouts=0 det_noise_sd=0.005 log_wandb=True optim_params.stepsize=0.15 spurious_prompt_id=4 datasets=['spurious_imagenet']
Exploration
conda activate dash_retrieval
CUDA_VISIBLE_DEVICES=... python src/3_sdxl_retrieval.py source_dir=output/2_optim_sdxl datasets=['spurious_imagenet'] model_id=paligemma
conda activate dash
CUDA_VISIBLE_DEVICES=... torchrun --nproc-per-node N --standalone src/X_vlm_eval.py source_dir=output/3_optim_sdxl_retrieval/ model_id=paligemma vlm_batchsize=4 datasets=['spurious_imagenet']
Exploitation
conda activate dash_retrieval
CUDA_VISIBLE_DEVICES=... python src/4_exploitation_retrieval.py source_dir=output/3_optim_sdxl_retrieval model_id=paligemma use_agnostic_folder=True
conda activate dash
CUDA_VISIBLE_DEVICES=... torchrun --nproc-per-node N --standalone src/X_vlm_eval.py source_dir=output/4_exploitation_prompt_retrieval/spurious_imagenet/llama_only_follow_up/prompt_4 model_id=paligemma
We provide JSON files containing URLs for all cluster of images resulting from DASH. These images were filtered by an object detector to ensure the absence of a given object but the source VLM responds "yes" to the question "Can you see a object in this image? Please answer only with yes or no."
Use the script src/download_images.py to download the images.
| VLM | #images | #clusters |
|---|---|---|
| PaliGemma-3B | 99.3K | 1892 |
| LLaVa-1.6-Vicuna | 162.4K | 3632 |
| LLaVa-1.6-Mistral | 78.5K | 2001 |
| VLM | #images | #clusters |
|---|---|---|
| PaliGemma-3B | 221.7K | 3895 |
| LLaVa-1.6-Vicuna | 252.0K | 4632 |
| LLaVa-1.6-Mistral | 133.8K | 3229 |
🚧 Under Construction 🚧 Stay tuned!
🚧 Under Construction 🚧 Stay tuned!
@article{augustin2025dash,
title={DASH: Detection and Assessment of Systematic Hallucinations of VLMs},
author={Augustin, Maximilian and Neuhaus, Yannic and Hein, Matthias},
journal={arXiv preprint arXiv:2503.23573},
year={2025}
}