NuclearDiffusion Creates Accurate Nuclear Energy Images via Fine-tuning

Generative AI, particularly text-to-image synthesis, has seen rapid advancements, but its application in highly specialized engineering domains like nuclear energy remains largely unexplored. General-purpose foundation models often produce inaccurate or conceptually inconsistent images when prompted with nuclear engineering concepts due to their lack of domain-specific knowledge. This limitation highlights the need for tailored solutions in such technical fields. A new study introduces NuclearDiffusion, a systematic approach to domain adaptation for nuclear text-to-image generation. The researchers curated a dataset of 1,000 captioned images covering various nuclear energy topics, including reactors, fuel cycles, and radiation. This dataset was then used to fine-tune three open-source diffusion models: Stable Diffusion XL (SDXL), SD-v3.5-Medium, and Flux.1. Evaluation showed that fine-tuning substantially improved SDXL's fidelity, offered limited gains for SD-v3.5-Medium, and no measurable improvement for Flux.1, indicating that the effectiveness of adaptation is highly dependent on the underlying generative architecture. When compared to leading commercial systems like GPT-Image-2, Gemini-3.1-Flash-Image, and Midjourney, the fine-tuned open-source models, particularly SDXL, produced more accurate and technically consistent outputs for specialized nuclear engineering prompts, even though commercial models performed well on broader concepts. This research establishes domain-specific fine-tuning as a viable method for creating reliable generative AI tools for niche applications.