FiSeR-DINOv3-ViT-L16

FiSeR DINOv3 ViT-L/16

This model is the image encoder trained by FiSeR with hierarchical supervised contrastive learning. It outputs a 1024-dimensional CLS representation. The released model is a backbone; attach a kNN, prototype, linear, or SVM head for the binary natural/synthetic decision described in the paper.

The released weights correspond to the wild_model19.pth checkpoint from zero-indexed epoch 19 (the 20th and final training epoch). The training-only source-center parameter is omitted from this backbone export.

Paper: arXiv:2606.00606
Code: heyongxin233/FiSeR

Usage

from transformers import AutoImageProcessor, AutoModel
import torch

model_id = "heyongxin233/FiSeR-DINOv3-ViT-L16"
processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id)
image = ...  # PIL.Image.Image
inputs = processor(images=image, return_tensors="pt")
with torch.inference_mode():
    features = model(**inputs).last_hidden_state[:, 0]

Use the FiSeR repository's evaluate.py to reproduce the paper's GPU-FAISS retrieval protocol. The published safetensors SHA-256 is 3bfda6c3040fe22f8fe82588bd4c8506572a6844a04c15bb5c8feebbbd9e11a5.

The fixed-layer direct-transfer average is 98.23 AUROC / 93.43 TPR@5% FPR over WildFake, Community Forensics, AIGIBench, Chameleon, and GenImage. Refer to the repository for dataset revisions and the exact per-dataset protocol. This checkpoint must not be used as a substitute for content provenance, safety review, or human moderation.

Training data and limitations

The model was trained on the WildFake training split. It can be sensitive to dataset composition, image post-processing, and generator families absent from WildFake. Dataset and DINOv3 licensing terms remain applicable.