6M-image study maps open-source AI art ecosystem with 154K LoRA models

Researchers from multiple institutions, led by Yiluo Wei, conducted the first large-scale empirical study of creator practices within the open-source image generation ecosystem. They constructed a novel dataset of 6 million images, each carrying detailed generation metadata—the precise recipe of models, prompts, and LoRA adapters used. By linking usage data from 22.4K base models and 154K LoRA models, the study maps how creators curate and combine community-contributed components. This contrasts sharply with closed-source tools like Midjourney, where workflows are opaque. The paper, accepted at ACM MM'26, provides insights into both the strengths (e.g., flexibility, fine-grained control) and obstacles (e.g., model discoverability, dependency management) of the ecosystem.

The authors make the dataset publicly available to serve as a practical reference for creators seeking to produce better artworks and for researchers studying generative AI workflows. Key findings highlight the diverse combinatorial patterns—some creators rely on a few popular base models while others experiment with many niche LoRAs—and the challenges of versioning and compatibility across the rapidly evolving model landscape. This work is a critical step toward making open-source image generation more sustainable and innovative, offering data-driven guidance for platform designers, model developers, and artists navigating the complex open-source model ecosystem.