Luma Photon Flash is a high-speed text-to-image foundation model developed by Luma Labs. Released alongside the standard Photon model, it is optimized for rapid inference and cost efficiency, targeting professional workflows that require high-volume generation and fast iteration. The model is built on Luma's bespoke Universal Transformer architecture, which allows for significantly faster processing speeds and lower operational costs compared to standard diffusion-based models.
The model is particularly noted for its performance in instruction following and spatial composition. According to Luma Labs' internal evaluations, Photon Flash achieves high preference scores in these categories, effectively following complex, multi-part prompts and maintaining coherent spatial arrangements. It is engineered to minimize common "AI artifacts," producing photorealistic textures, nuanced lighting, and consistent results across various aspect ratios.
Key features of Photon Flash include a robust character consistency system and sharp text rendering. By utilizing a single input image as a character reference, the model can maintain a subject's identity across different scenes. It also excels at generating legible, correctly spelled typography within images. Additionally, the model supports style references and image-to-image workflows, allowing users to guide the aesthetic and structural output with a high degree of precision through natural language and visual examples.