Request for HP-Image-40K test split metadata and evaluation scripts

[LilaKen]

Hi HiFi-Inpaint authors,

Thanks for releasing the code, dataset, and checkpoints — great work!

I am trying to reproduce the quantitative results in Table 1 of the paper. The paper says that 1,000 samples were split from HP-Image-40K for evaluation, were not used for training, and all methods were evaluated at 1024 × 576 with CLIP-T, CLIP-I, DINO, SSIM, SSIM-HF, LAION-Aes, and Q-Align-IQ.

However, the current public HP-Image-40K release appears to contain only the training metadata/files:

• meta_data.json with 43,632 samples
• ref_images.zip
• gt_images.zip
• condition_images.zip
• masks.zip

I could not find a released test.json / val.json / split file in the Hugging Face dataset, and I also could not find an official evaluation script in this repository.

Could you please share one of the following, if possible?

1. The exact 1,000-sample evaluation split metadata, or at least the list of sample indices/filenames used for Table 1.
2. The official evaluation scripts/configs for CLIP-T, CLIP-I, DINO, SSIM, SSIM-HF, LAION-Aes, and Q-Align-IQ.
3. Any details about whether the released LoRA checkpoint was trained exactly with the same training data described in the paper, i.e. HP-Image-40K plus the internal ~14K dataset.

This would help make the reported results much easier to reproduce faithfully.

Thank you!