5 distinct image sources across real and synthetic categories provide training data for forgery detection, specifically utilizing DiffusionDB and LAION-Aesthetics. Evaluation sets derived from Midjourney, PixArt-alpha, and GPT-4o allow for testing cross-generator generalization.
Use Cases
- Train a binary classifier to distinguish between real and fake images using the DiffusionDB and LAION-Aesthetics training splits.
- Benchmark the detection accuracy of models on unseen generative architectures using the PixArt-alpha evaluation set.
- Analyze the visual artifacts of GPT-4o generated images compared to real-world LAION-Aesthetics samples.
Strengths
- Training set utilizes fake images from DiffusionDB and real images from LAION-Aesthetics.
- Evaluation sets include Midjourney images sourced from the GenImage dataset.
- Includes synthetic images from GPT-4o via the ShareGPT-4o collection.
- Features PixArt-alpha samples licensed under OpenRail for testing model diversity.