dhryou/VHA-plus

VHA+

VHA+ is a lightweight hallucination assessor for ground-truth and super-resolved image pairs. It uses a shared DINOv3 ViT-B/16 encoder and predicts VHA scores from image-pair features. Higher scores indicate more hallucination relative to the reference.

• Paper
• Code and installation
• Project page
• VHA-bench

Download and evaluate

This repository provides the original step-20,000 checkpoint as vhaplus.pt, without optimizer state. Use it with the accompanying GitHub implementation; it is a custom PyTorch checkpoint.

After installing the GitHub repository and preparing VHA-bench under data/benchmark/:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="dhryou/VHA-plus",
    filename="vhaplus.pt",
    local_dir="data/checkpoints",
)

python scripts/vha+/evaluate.py \
  --checkpoint data/checkpoints/vhaplus.pt \
  --output outputs/vhaplus_evaluation

Checkpoint details

• Shared DINOv3 ViT-B/16 encoder; the last four transformer blocks are trained.
• Final normalized CLS features and the checkpoint's stored reporting scale.
• Input resolution: 512 × 512.
• Training step: 20,000.
• The config.json records the model configuration and reporting scale; the GitHub implementation loads vhaplus.pt directly.

On VHA-bench, the released model achieves Spearman 0.659 and Pearson 0.675 with human hallucination judgments. These are the paper's fixed benchmark results, not a new evaluation performed during this upload.