Developing an Image Generation Model for Creative Production Using Rights-Cleared Images

This project is the result of my work at the 2026 summer internship at Preferred Networks, Inc. (PFN). Using only images whose copyrights have expired or whose rights have been waived, I trained a text-to-image model from scratch. Because the origin of training data is often questioned in creative production, the goal was to develop an image generation model trained on rights-cleared images that can be used in actual production work.

The training data was PD12M, about 11.4 million public-domain images, and I trained a model with about 600M parameters. The model trained only on PD12M had two problems: it was weak at depicting indoor scenes, and its outputs tended to look like black-and-white photos or prints.

For the first problem, I generated indoor images with Infinigen, which procedurally generates 3D scenes, and added them to the training data. For the second, I post-trained the model on modern photos. As a result, the indoor instruction-following accuracy for the trained room types rose from 41.4% to 58.3%, on par with existing models. I also prototyped style transfer and layout control as features for production use.

Details are published in the PFN tech blog article "Developing an Image Generation Model for Creative Production Using Rights-Cleared Images" (in Japanese).