Luma's unified reasoning model generates and edits images in a single autoregressive pass. 2 credits to generate, 2 or 3 credits to edit, five aspect ratios, no watermark, on every Aura AI plan.
Luma Uni-1 is the image model Luma AI announced in March 2026, alongside its Luma Agents product, and the company's first "unified understanding and generation" model. Luma describes it as a multimodal reasoning model that can generate pixels. Where Dream Machine and Ray 2 made Luma known for video, Uni-1 is the company's bid for stills: one model that understands a request, plans the picture and produces it, then edits it when you ask for changes.
The architecture is the interesting part. Uni-1 drops the diffusion pipeline used by most image generators in favor of a decoder-only autoregressive transformer. Text tokens and image tokens sit in one interleaved sequence, so inference and pixel generation happen in a single forward pass, much like a language model writing a reply. That design is why Uni-1 can follow layered directions about camera angle, lighting and color, accept up to 9 reference images upstream to lock identity and composition, and apply plain-English edits such as swapping a background or shifting the light. Luma reports 2048 pixel output upstream and, on its own human preference evaluation at launch, first place in overall, style and editing, and reference-based generation, with second place in text-to-image.
On Aura AI, Uni-1 appears in three forms. Luma Uni-1 generates from text for 2 credits per image in five aspect ratios. In the image editor, Luma Uni-1 Edit costs 2 credits and Luma Uni-1 Edit Max costs 3 credits per edit. All three are available on every plan. Rivals in the same unified category include Nano Banana, GPT Image 2 and FLUX 2, and you can switch between them without leaving the tool.
Six things a reasoning-first image model does differently
Understanding and generation share a single autoregressive transformer. The model reads your intent, plans the layout and writes the pixels in the same sequence, with no separate diffusion stage.
Upload an image and describe the change: "make it golden hour", "remove the sign", "put her in a red coat". Uni-1 edits the region you mean and leaves the rest intact. Two tiers on Aura AI, Edit and Edit Max.
Upstream Uni-1 accepts up to 9 reference images to preserve identity, pose and composition. Luma reports first place in reference-based generation on its human preference evaluation.
Camera angles, lighting setups, color grades and even rough doodles or collages can steer the output. The model treats these as instructions to reason about rather than loose style hints.
Launch coverage cites 76 visual styles, from realistic photography to comics and ukiyo-e. Uni-1 ranked first for style and editing in Luma's Elo comparison, so style changes on existing images are a strength.
For complex prompts the model plans the arrangement of elements and can add annotations, which makes it usable for storyboards, marketing layouts and diagrams, not only single-subject shots.
Four steps from prompt to an edited, downloadable image
Flat cost per image, the same in every aspect ratio
| Variant | Mode | Credits per image | Best for |
|---|---|---|---|
| Luma Uni-1 | Text to image | 2 credits | Directed scenes, storyboards, styled stills |
| Luma Uni-1 Edit | Image edit | 2 credits | Background swaps, relighting, local fixes |
| Luma Uni-1 Edit Max | Image edit | 3 credits | High-fidelity edits on products and portraits |
How Luma's unified model compares to other generate-and-edit models on Aura AI
| Feature | Luma Uni-1 | Nano Banana Pro | FLUX 2 | GPT Image 2 |
|---|---|---|---|---|
| Maker | Luma AI | Black Forest Labs | OpenAI | |
| Text-to-image credits | 2 | 3 | 2 | 4 |
| Image edit credits | 2 (Edit Max 3) | 2 | 2 | 4 |
| Architecture | Autoregressive, unified | Gemini-based, unified | Diffusion transformer | Unified multimodal |
| Reference images upstream | ✓ Up to 9 | ✓ Multi-image | ✓ Multi-reference | ✓ Multi-image |
| Edit tiers on Aura AI | 2 | 1 | 3 (Edit, Pro Edit, Max Edit) | 1 |
| Aspect ratios on Aura AI | 5 | 5 | 5 | 5 |
| Plan | All plans | All plans | All plans | All plans |
Where reasoning and editing in one model pay off
Take one packshot and produce a whole catalog: new backdrop, new light, same product. Use Edit Max when the label and materials must stay exact.
Describe shot, lens and blocking and Uni-1 plans the composition. Generate the sequence at 2 credits per frame, then animate the keeper in Image to Video.
Change an outfit, fix the lighting or move a subject to a new setting with a sentence. The model keeps identity and pose while applying the change.
Hero images with planned negative space for copy, or annotated explainers where the model places labels where they belong.
Common questions about Luma Uni-1 on Aura AI
Luma's unified reasoning model, 2 credits to generate, 2 or 3 to edit, five aspect ratios, no watermark. Available on every plan, worldwide.