nakas/gpttrace-img-v2

GPTTrace image v2 training data

Training data for GPTTrace's AI-generated image detectors. Everything here comes from permissively licensed sources or was generated by us from CC0 photos. Each part keeps its source licence; check the source before reuse.

All files are parquet shards with these columns: id, src, gen, kind, label, w, h, jpeg, image. The image column holds the original bytes.

Folder / prefix	Content	Source	Licence
vae/vae-*	Real photos round-tripped through open autoencoders. Six VAEs: SD, SDXL, FLUX.1-style, FLUX.2, Qwen-Image, Sana. Also includes partial "local" edits. The JPEG settings match each real twin.	bitmind/megalith-small photos (CC0); our generation	CC0
ext/dragon-*	25 diffusion models	lesc-unifi/dragon	CC BY-SA 4.0
ext/elsa-*	SD-family	elsaEU/ELSA1M_track1	CC BY 4.0
ext/oip-*	FLUX-dev, SD3.5 Large	data-is-better-together/open-image-preferences-v1	Apache-2.0
ext/nano-*	Gemini 2.5 Flash Image	bitmind/nano-banana	MIT
ext/gpt4o-*	GPT-4o image	FreedomIntelligence/ShareGPT-4o-Image	Apache-2.0
ext/ddb-*	SD 1.x	poloclub/diffusiondb	CC0
ext/rapid-*	Generators from 2023 to 2026, sampled evenly per model	Rapidata/text-to-image-human-preference-6M	CC BY 4.0
t2i/*	Z-Image-Turbo, FLUX.2-klein, Sana generations	our generation (Apache-2.0 models)	CC0

Attribution: thanks to the creators of the datasets above. Rapidata data © Rapidata, CC BY 4.0. DRAGON © the DRAGON authors, CC BY-SA 4.0. ELSA © the ELSA project, CC BY 4.0.

No benchmark or evaluation data is included. Benchmark images and near-duplicates of them are excluded from training (see dedupe/).