Fuser Apps are here 🚀
Fuser Apps are here. Free generations for the next month 💫
Let's GobyLuma Labs
Compose multi-character narratives, complex storyboard panels, and authentic manga layouts with physically grounded spatial reasoning
Uni-1 replaces random diffusion noise with an autoregressive planning engine that reasons through spatial hierarchy, reference tags, and lighting physics before generating final imagery.
Explore how autoregressive transformer architecture, eight-slot reference anchoring, and native manga screentone rendering elevate visual storytelling.
A curated gallery demonstrating physically calculated lighting, complex multi-character staging, and bold editorial palettes created with Uni-1.
From mangakas drafting serialized pages to industrial designers staging luxury prototypes, see how creators leverage agentic image generation.
Luma Uni-1 uses a decoder-only autoregressive transformer that plans scenes before rendering, rather than denoising random pixel noise like traditional diffusion models. This agentic reasoning phase calculates spatial geometry, lighting vectors, and prompt constraints first, resulting in physically plausible depth, realistic atmospheric scattering, and superior multi-character consistency.
You can upload up to eight reference images and bind them by referencing tags like IMAGE1 (CHARACTER), IMAGE2 (STYLE), or IMAGE3 (COMPOSITION) directly in your text prompt. Explicitly linking the images in your prompt text instructs the model's planning engine exactly how to map character traits, lighting setups, or artistic styles onto the final composition.
Manga mode is a dedicated style preset that generates authentic manga page layouts with clean pen lines, traditional screentones, and integrated typography. It is designed to work natively with portrait aspect ratios (such as 2:3, 9:16, 1:2, and 1:3) to replicate classic Japanese ink-and-wash publication formats.
No, Luma Uni-1 does not support negative prompting. Because it is an autoregressive vision-language model that parses natural instructions, negative keywords like 'no blur' or 'ugly' will be interpreted literally as visual subjects. Instead, describe what you want to see using positive, descriptive natural language or structured Markdown directives.
Enable Web Search when you need accurate depictions of contemporary branding, verified architectural landmarks, real-world products, or historical cultural artifacts. The planning agent consults live web context to verify visual details before rendering, though it is best left disabled for fantasy or abstract scenes to avoid extra generation latency.
Compose multi-character narratives, complex storyboard panels, and authentic manga layouts with physically grounded spatial reasoning