Synthetic Image Forensics

Complete model weights for synthetic-image classification, artifact localization and explanation generation. Install the code and dependencies from GitHub.

CUDA_VISIBLE_DEVICES=0 python scripts/infer.py \
  --model broken-ice/synthetic-image-forensics \
  --image path/to/image.png \
  --output outputs/predictions

Each prediction contains the image class, an explanation and the final artifact mask. Images predicted as real receive an empty artifact mask. Use a new output directory for each run. For multiple images, replace --image ... with --manifest data/images.jsonl. Manifest records contain sample_id and image_path; see the GitHub README for data formats, evaluation and training commands.

The model downloads openai/clip-vit-large-patch14-336 automatically at the revision specified in config.json. CLIP is cached separately and is not included in these weight shards. All other model components are included; no base-model or adapter checkpoint assembly is required.

--model also accepts a downloaded model directory. Once both the model and CLIP are cached, use --local-files-only to run offline.

See LICENSE and NOTICE for licensing and attribution.

Downloads last month
83
Safetensors
Model size
8B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support