How Many Images Does It Take? Estimating Imitation Thresholds in Text-to-Image Models
text-to-imagecopyrightimitationgenerative-aitraining-data
Abstraction: Empirical threshold at which text-to-image models begin imitating training concepts
Key points:
- Imitation threshold: the minimum training instances of a concept needed for a model to generate recognizably similar output — estimated at 200–700 images depending on domain and model
- Studied two domains: human faces and art styles; evaluated four text-to-image models on three pretraining datasets
- Proposes an efficient estimation approach that avoids training models from scratch for each test
- Provides an empirical basis for copyright violation claims and a guiding principle for model developers aiming to comply with privacy and copyright law
- Code available at github.com/vsahil/MIMETIC-2
Connections: Text To Image · Generative AI · Copyright AI · Training Data
Source: https://arxiv.org/abs/2410.15002