How Many Images Does It Take? Estimating Imitation Thresholds in Text-to-Image Models

TL;DR

This work defines the imitation threshold as the minimal training exposure needed for a text-to-image model to imitate a concept. It introduces MIMETIC2, an observational, cost-efficient framework that combines concept-frequency estimation, domain-specific imitation scoring, and change-point detection to estimate thresholds without retraining models. Across two domains—human faces and art styles—and four text-to-image models trained on three datasets, it finds thresholds typically in the 200–700 image range and highly dependent on domain and model. Human-subject validation demonstrates strong alignment with automatic scores, supporting the thresholds' relevance for copyright/privacy considerations and policy guidance. Overall, the approach provides a principled, scalable method to bound imitation risk and inform responsible development of text-to-image systems.

Abstract

Text-to-image models are trained using large datasets of image-text pairs collected from the internet. These datasets often include copyrighted and private images. Training models on such datasets enables them to generate images that might violate copyright laws and individual privacy. This phenomenon is termed imitation -- generation of images with content that has recognizable similarity to its training images. In this work we estimate the point at which a model was trained on enough instances of a concept to be able to imitate it -- the imitation threshold. We posit this question as a new problem and propose an efficient approach that estimates the imitation threshold without incurring the colossal cost of training these models from scratch. We experiment with two domains -- human faces and art styles, and evaluate four text-to-image models that were trained on three pretraining datasets. We estimate the imitation threshold of these models to be in the range of 200-700 images, depending on the domain and the model. The imitation threshold provides an empirical basis for copyright violation claims and acts as a guiding principle for text-to-image model developers that aim to comply with copyright and privacy laws. Website: https://how-many-van-goghs-does-it-take.github.io/. Code: https://github.com/vsahil/MIMETIC-2.

Figures (40)

• Figure 1: An overview of the imitation phenomenon where we seek the imitation threshold -- the point at which a model was exposed to enough instances of a concept that it can reliably imitate it. The figure shows four concepts (e.g., Van Gogh's art style) that have different counts in the training data (e.g., 213K for Van Gogh). As the image count of a concept increases, the ability of the text-to-image model to imitate it increases (e.g., Piet Mondrian's and Van Gogh's art styles have higher imitation). The imitation threshold represent the number of instances a model has to be trained on such that humans recognize such concept in generated images.
• Figure 2: Overview of MIMETIC2's methodology to estimate the imitation threshold. In Step 1, we estimate the frequency of each concept (belonging to a domain) in the pretraining data by obtaining the images that contain the concept of interest. In Step 2, we use the filtered images of each concept (obtained in Step 1) and compare them to the generated images to measure imitation (using $g$ that receives training and generated images). We repeat this process for each concept to generate the imitation score graph, and then determine the imitation threshold with a change detection algorithm.
• Figure 3: LAION captions that mention 'Mary Lee Pfeiffer', the mother of Tom Cruise. She is not always present in the images (the rightmost image).
• Figure 4: Examples of real celebrity images (top) and generated images from a text-to-image model (bottom) with increasing image counts from left to right (3, 273, 3K, 10K, and 90K, respectively). The prompt is "a photorealistic close-up image of {name}".
• Figure 5: Human Face and Art Style imitation graphs for SD1.1 using the Celebrities and Classical art style sets. The x-axis is the image frequencies and the y-axis is the imitation score averaged over the five prompts. Concepts with zero image frequencies are shaded in light gray. We show the mean imitation score and its variance over the five image generation prompts for each concept. The red vertical line indicates the imitation threshold found by the change detection algorithm, and the horizontal green line represents the average imitation scores before and after the threshold.