eerie-road/rorem-discriminator

RORem discriminator — safetensors mirror

The RORem discriminator: a classifier over the SDXL-Inpainting UNet (a rank-4 LoRA plus a cls_pred_branch head on the down/mid blocks) that scores whether an object has been removed from an image. Released by the RORem authors as a by-product of their human-in-the-loop training; upstream it is distributed from a Google Drive folder linked in the project README, with no Hugging Face mirror of its own. Hosted here so that sayvoy-look can pull it through a pickle-free, digest-pinned path.

Field	Value
Paper	RORem: Training a Robust Object Remover with Human-in-the-Loop — Ruibin Li, Tao Yang, Song Guo, Lei Zhang. CVPR 2025.
Upstream	https://github.com/leeruibin/RORem (weights: the "RORem-Discriminator" Google Drive folder linked from its README)
Base model	diffusers/stable-diffusion-xl-1.0-inpainting-0.1
License	Apache-2.0, as declared in the upstream README ("This project is released under the Apache 2.0 license"). See the note below.

Files

File	Bytes	SHA-256	Origin
pytorch_lora_weights.safetensors	11,483,632	a3ae1c8f3d0a446fe403998f5c732a6c0d51939c92e5d5c47195875e3b4b8967	the authors' own safetensors, unchanged
cls_pred_branch.safetensors	157,313,466	3d9d52b12bdbed0cc8eb287d5d43c7b5d1ab6f3eb647194257463cf930408842	converted from the authors' cls_pred_branch.pt (see below)
empty_prompt_embeds.safetensors	636,080	2446465d79e9456d3f27b8f0073614a9dece0588dbd7d1971e64c10d83f3373e	derived: the empty prompt encoded by SDXL-Inpainting 0.1's two CLIP text encoders

cls_pred_branch.pt upstream is a PyTorch zip whose payload is a pickle. It was scanned statically with pickletools (nothing executed; the only globals are collections.OrderedDict, torch.HalfStorage, torch._utils._rebuild_tensor_v2 — a plain dict of 14 fp16 tensors), loaded once under the restricted unpickler, written out as safetensors, read back and compared bit for bit, then pinned.

empty_prompt_embeds.safetensors is not an upstream file: the discriminator is run with an empty prompt, and these are that prompt's prompt_embeds / pooled_prompt_embeds from the base model's text encoders, cached so the judge runs without loading them.

How it is used

The discriminator runs at its training resolution (512) over the SDXL-Inpainting 0.1 UNet with the LoRA applied and the head on the down/mid-block features. It measures removal, not fill quality: in our evaluation its score separates "object gone" from "object still there" well (AUC ≈ 0.98) and is uncorrelated with the fidelity of the fill. The model definition (unet_sdxl_discriminator.py, itself a derivative of diffusers' UNet2DConditionModel) lives in the consumer, not in this repo.

License note

The upstream README declares Apache-2.0 and links a LICENSE file that does not exist in the repository. This mirror relies on the README's declaration and ships the standard Apache-2.0 text as LICENSE. Credit belongs to the RORem authors; please cite the paper.

@inproceedings{li2025rorem,
  title     = {RORem: Training a Robust Object Remover with Human-in-the-Loop},
  author    = {Li, Ruibin and Yang, Tao and Guo, Song and Zhang, Lei},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025}
}