Instructions to use aiorscam/pgc with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use aiorscam/pgc with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-classification", model="aiorscam/pgc") pipe("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/parrots.png")# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("aiorscam/pgc", dtype="auto") - Notebooks
- Google Colab
- Kaggle
| library_name: transformers | |
| tags: | |
| - image-classification | |
| - ai-generated-image-detection | |
| - dinov2 | |
| - pgc | |
| license: other | |
| # AiorScam PGC Deployment Mirror | |
| This repository mirrors the assets needed by the `aiorscam/pgc-inference` | |
| dedicated endpoint implementation. | |
| ## Sources | |
| - PGC source repository: https://github.com/xiaoyu6868/PGC | |
| - PGC upstream commit used for parity review: | |
| `e049d1ed97cc0ed066b61d445a7cced216bf550c` | |
| - PGC checkpoints source: https://modelscope.cn/models/xiaoyuzhou68/PGC_ckpt | |
| - DINOv2 backbone source: https://huggingface.co/facebook/dinov2-large | |
| ## Layout | |
| ```text | |
| checkpoints/ | |
| PGC_train_sdv1_4_ckpt.pth | |
| PGC_train_progan_ckpt.pth | |
| PGC_train_progan_sdv1_4_ckpt.pth | |
| dinov2-large/ | |
| config.json | |
| model.safetensors | |
| preprocessor_config.json | |
| ``` | |
| ## Endpoint Configuration | |
| The default endpoint variant is the joint PGC checkpoint: | |
| ```text | |
| PGC_VARIANT=progan_sdv1_4 | |
| PGC_CHECKPOINT_DIR=/repository/checkpoints | |
| PGC_DINO_PRETRAINED_ROOT=/repository | |
| PGC_FAKE_THRESHOLD=0.5 | |
| PGC_REAL_THRESHOLD=0.5 | |
| ``` | |
| Supported variants are `sdv1_4`, `progan`, and `progan_sdv1_4`. | |
| ## Notes | |
| PGC is a promptless binary real/synthetic image detector. The deployed endpoint | |
| does not send or synthesize text prompts; parity with the author implementation | |
| comes from the DINOv2 preprocessing path, checkpoint selection, and binary | |
| decision threshold. | |