NuclearDiffusion: Fine-Tuning Diffusion Models for Nuclear Energy Imagery

A new arXiv preprint introduces NuclearDiffusion, a systematic study on adapting text-to-image diffusion models for the specialized domain of nuclear engineering. The research addresses a critical gap: general-purpose AI models often generate physically incorrect or conceptually inconsistent images in this field due to a lack of domain-specific knowledge. The authors curated a dataset of 1,000 captioned nuclear energy images covering reactors, fuel cycles, radiation, and related concepts. They fine-tuned three state-of-the-art open-source models—Stable Diffusion XL (SDXL), SD-v3.5-Medium, and the flow-matching Flux.1 model—on this dataset. Performance was evaluated using quantitative image-similarity metrics and qualitative expert assessment, comparing the fine-tuned models against zero-shot baselines. The study represents one of the first attempts to adapt diffusion models for nuclear text-to-image generation, potentially enabling more accurate visualizations for education, communication, and research in nuclear engineering. The paper is available on arXiv under the identifier 2608.04030.