You can deploy this model and start earning money today!

QinmingZhou/
OSOR

Name: OSOR
Availability: InStock

Views

No views yet

OSOR

Paper | Code

OSOR is a one-step diffusion inpainting framework for effect-aware object removal. This repository provides LoRA-based checkpoints for two implementations built with Diffusers and PEFT on FLUX.1 Fill [dev] and Stable Diffusion XL Inpainting.

Released Checkpoints

Model	Stage	Checkpoint	Base model	Model license
OSOR-FLUX-Fill	Phase I	`osor-fluxfill/weights/fluxfill_phase1.pth`	FLUX.1 Fill [dev]	FLUX.1-dev Non-Commercial License
OSOR-FLUX-Fill	Phase II	`osor-fluxfill/weights/fluxfill_phase2.pth`	FLUX.1 Fill [dev]	FLUX.1-dev Non-Commercial License
OSOR-SDXL-Inpainting	Phase I	`osor-sdxlinpainting/weights/sdxlinpainting_phase1.pth`	SDXL Inpainting 0.1	CreativeML Open RAIL++-M
OSOR-SDXL-Inpainting	Phase II	`osor-sdxlinpainting/weights/sdxlinpainting_phase2.pth`	SDXL Inpainting 0.1	CreativeML Open RAIL++-M

Phase I is trained with object-core supervision. Phase II introduces effect-aware supervision using CORNE.

Download

Download all released checkpoints:

hf download QinmingZhou/OSOR --local-dir ./OSOR-weights

Download one implementation only:

bash

hf download QinmingZhou/OSOR \
  --include "osor-fluxfill/weights/*" \
  --local-dir ./OSOR-weights

hf download QinmingZhou/OSOR \
  --include "osor-sdxlinpainting/weights/*" \
  --local-dir ./OSOR-weights

Usage

Inference and training instructions are provided in the corresponding directories of the OSOR code repository:

osor-fluxfill/
osor-sdxlinpainting/

The released .pth files contain the trained LoRA parameters and, depending on the implementation and training phase, additional trainable output-layer parameters. They are raw PyTorch state dictionaries produced by the OSOR training code rather than standalone Diffusers adapter directories. Load them with the corresponding OSOR implementation instead of calling load_lora_weights() directly.

Data

Training dataset: CORNE
Validation benchmark: CORNE-Val
Text-removal benchmark: TextEraseBench
Anime-removal benchmark: AnimeEraseBench

License

The two checkpoint families inherit different base-model licenses:

OSOR-FLUX-Fill is subject to the FLUX.1-dev Non-Commercial License and is not released for commercial use.
OSOR-SDXL-Inpainting is subject to the CreativeML Open RAIL++-M License, including its use-based restrictions.
The source code in the OSOR GitHub repository is released under the MIT License. The code license does not replace or override the licenses of the model weights or their base models.

Users are responsible for reviewing the applicable license before using or redistributing a checkpoint.

Citation

bibtex

@inproceedings{zhou2026osor,
  title     = {OSOR: One-Step Diffusion Inpainting for Effect-Aware Object Removal},
  author    = {Zhou, Qinming and Sun, Chenxi and Kong, Deyang and He, Junhao and Tang, Xiangheng and Yu, Peike and Wu, Haotian and Cao, Leilei and Zhang, Linfeng},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.28094}
}