NuclearDiffusion: Text-to-Image Foundation Models for Learning Nuclear Energy Concepts

Generative image models are getting better at broad prompts, but specialized engineering subjects still expose their weak spots. A new nuclear-energy study tackles that gap by fine-tuning open-source diffusion models on a curated set of 1,000 captioned images covering reactors, fuel cycles, radiation, and related concepts.

Three models were tested: Stable Diffusion XL, SD-v3.5-Medium, and Flux.1. The results were uneven. SDXL improved substantially after fine-tuning, SD-v3.5-Medium gained only a little, and Flux.1 showed no measurable improvement. That makes the work especially interesting for developers because it suggests domain adaptation is not just about throwing more data at a larger model; the underlying generative architecture can determine whether fine-tuning actually sticks.

The models were judged with both image-similarity metrics and expert review, then compared against GPT-Image-2, Gemini-3.1-Flash-Image, and Midjourney. The commercial systems produced convincing general nuclear imagery, but they still stumbled on specialized engineering prompts where the fine-tuned open-source models were more technically consistent.

For game teams, the practical lesson is broader than nuclear engineering. If you need concept art, UI mockups, or training visuals for a niche setting, a small, well-curated dataset plus targeted fine-tuning may beat a general-purpose model on accuracy and consistency. The study also reinforces that trustworthy generative tools for domain work will likely come from adaptation pipelines, not one-size-fits-all foundation models.