---
base_model: facebook/dinov3-vitl16-pretrain-lvd1689m
datasets:
- Jason37437/Doc-Protocol-Data
language:
- en
library_name: pytorch
license: mit
metrics:
- f1
pipeline_tag: image-segmentation
tags:
- image-forgery-localization
- document-forensics
- computer-vision
---

# RiSE

Checkpoints for **RiSE: Residual Subspace Expert for Generalizable Text-Centric Image Forgery Localization**.

Paper: [RiSE: Residual Subspace Expert for Generalizable Text-Centric Image Forgery Localization](https://huggingface.co/papers/2610.24480)

Code and usage: https://github.com/KahimWong/RiSE

| Model | Checkpoint | Configuration |
|---|---|---|
| RiSE-1 | [RiSE-1.pth](RiSE-1.pth) | [RiSE-1.json](RiSE-1.json) |
| RiSE-3 | [RiSE-3.pth](RiSE-3.pth) | [RiSE-3.json](RiSE-3.json) |
| RiSE-9 | [RiSE-9.pth](RiSE-9.pth) | [RiSE-9.json](RiSE-9.json) |
| RiSE-9-DCT | [RiSE-9-DCT.pth](RiSE-9-DCT.pth) | [RiSE-9-DCT.json](RiSE-9-DCT.json) |

The original checkpoint bytes are preserved; only public filenames are simplified. SHA-256 hashes and sizes are in `checksums.json`. Load these trusted PyTorch checkpoints using the strict loader in the code repository.

All models use DINOv3 ViT-L/16. RiSE-9-DCT uses the 1024-channel DCT projector and fusion at zero-based block index 10. No legacy DCT checkpoints are included. Inference accepts 512x512 images; DCT inference requires JPEG input. Pretrained-backbone terms apply separately.

## Single-Image Inference

Install the [RiSE code](https://github.com/KahimWong/RiSE) and download a checkpoint with its matching config:

```bash
git clone https://github.com/KahimWong/RiSE.git
cd RiSE
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

hf download Jason37437/RiSE RiSE-9.pth RiSE-9.json --local-dir checkpoints
python infer.py \
  --checkpoint checkpoints/RiSE-9.pth --config checkpoints/RiSE-9.json \
  --image /path/to/image.jpg --output outputs/demo --gh-order 5
```

## Doc Protocol Benchmark

Train on DocTamper and evaluate on seven test sets. In-domain FCD/SCD/Test images are compressed exactly once using the official final quality-factor record (`record[-1]`). Cross-domain T-SROIE/OSTF/TPIC-13/RTM images are original RGB JPEG patches without additional resizing or compression.

![Table 2: Doc Protocol F1 benchmark](table2.svg)

## Seven-Set Evaluation

After preparing `configs/data.json` as described in the [code README](https://github.com/KahimWong/RiSE#-data-preparation), run:

```bash
python evaluate.py \
  --checkpoint checkpoints/RiSE-9.pth --config checkpoints/RiSE-9.json \
  --data configs/data.json --output outputs/rise9_eval --gh-order 5
```