5 frozen ResNet50 expert classifiers and 4 MoE gating network checkpoints for forensic detection and attribution of AI-generated images.
Trained as part of the
unmasking-synthetic-images project — see the GitHub repo for the full training pipeline, evaluation
suite, and live demo.
Each expert is a ResNet50 (pretrained on ImageNet1K) fine-tuned end-to-end as a binary classifier (real vs synthetic) on its own Stable Diffusion variant.
Trained on RunPod (RTX 4000 Ada, 20GB).
Each gating network combines the 5 frozen experts. Only the gating module is trained; experts remain frozen.
Trained on RunPod (RTX PRO 6000 Blackwell, 96GB), batch_size=512, BF16 mixed precision.
The live demo and Docker setup pull these checkpoints automatically. To load manually:
1from huggingface_hub import hf_hub_download
2
3REPO = "enricoroncuzzi/unmasking-synthetic-images-models"
4
5# Load one expert
6expert_ckpt = hf_hub_download(REPO, "experts/sd15/best-epoch=30-val_loss=0.0006-val_acc=1.0000.ckpt")
7
8# Load one MoE gating network
9moe_ckpt = hf_hub_download(REPO, "moe/logit/best-epoch=51-val_loss=0.0994-val_acc=0.9590.ckpt")
