Luma Uni-1 updates reference-guided image generation with sketch and multi-input controls

• Luma has opened Uni-1 as a reference-led image model, with Luma's launch post framing it around intent understanding rather than longer prompt engineering.
• The clearest new workflow in Luma's control demo is multi-input direction: creators can combine faces, environments, action poses, and rough sketches into one composed output.
• Luma's style demo is also positioning Uni-1 as taste-aware, showing outputs that move between editorial portraiture, manga panels, meme imagery, and lo-fi surreal scenes.
• Early practitioner examples in DreamLabLA's clip suggest Uni-1 is already being used upstream in production, from keyframe generation to faster material-map and asset work.

Uni-1 is now live as an image model on Luma's site, with the product page describing it as a multimodal system for image generation, editing, and reference-grounded control. Luma's public rollout centers on three claims: the model can infer intent, follow direction from examples instead of dense prompt syntax, and handle culturally specific visual styles with fewer misses.

The most concrete change for creators is how much weight references appear to carry. In Luma's examples, two portraits become a single cinematic saloon scene; a crystal bust plus a luxury interior become a full-body fashion image in the same material language; and two faces plus a sword-fight frame become a staged medieval duel. Another example turns annotated concept art into polished character renders, which is a stronger promise than simple style transfer.

The rollout examples imply two practical use cases. First, Uni-1 looks built for art direction by assembly: bring separate references for subject, setting, pose, and finish, then let the model reconcile them. Second, it looks useful for sketch-to-image and look-dev work, where the input is less a finished prompt than a packet of visual constraints.

That lines up with early usage from DreamLabLA's network, where Uni-1-made keyframes were turned into a finished short Liminal Memories. The same thread context says the model was also used to generate consistent PBR, normal, and displacement maps, then paired with triplanar and image projection workflows to speed asset creation without dropping quality.