Luma Uni-1 combines visual understanding and image generation in a model accessed through Luma's hosted application and image API. To guide a new image, provide references and state each one's role; the Agents API accepts up to nine image_ref entries, while editing an existing image uses type: "image_edit" with source. Technical overview, Product guide, Generation contract
What are the key facts about Luma Uni-1?
| Field | Verified information |
|---|---|
| Developer | Luma. Technical overview |
| Released | Launched in 2026; an exact day is not published in the cited company summary. Company announcement |
| Type | Decoder-only autoregressive transformer with interleaved text and image representations. Architecture |
| Size or parameters | Not published in the cited technical overview. Technical overview |
| License and access | Hosted Luma product and API access; no open-weight release documented. Product guide, API guide |
| Where it runs | Luma-hosted inference accessed through the application or API. Product guide, API documentation |
How do reference images guide Luma Uni-1 generation?
Luma describes Uni-1 as combining instruction analysis, constraint resolution, and composition planning with generation.
Its technical overview demonstrates scene transformations, reference-based direction, and image sequences built around a consistent setting. These are vendor demonstrations rather than independent guarantees. Technical report
The useful prompting implication is to describe relationships, not just objects. For a room scene, specify which object sits beside the window, which direction the viewer faces, and which parts of the layout must remain unchanged.
Try a brief with a clear spatial requirement: a reading chair to the left of a floor lamp, a small table between the chair and a window, and warm light falling across the table. Judge those relationships before judging the decorative style.
For reference-guided work, Luma documents using multiple images to inform the result.
Its generation guide recommends naming each reference's role, such as composition or color palette, so the requested contribution is explicit. Reference instructions
A suggested test is to pair a room-layout reference with a separate material reference. Ask for the layout of the first and the fabric treatment of the second. Make your acceptance checklist identify which features come from each input.
Inspect preservation and transformation separately. First compare the arrangement of objects with your layout reference; then inspect fabric, color, and lighting.
This gives you a clearer reason to accept an image than a general impression of similarity.
What are Luma Uni-1 reference and output limits?
Luma's technical overview describes a decoder-only autoregressive transformer but does not publish its parameter count. The product materials document hosted access and provide no open-weight download. Architecture, Product guide
The Agents API permits up to nine image_ref entries. With type: "image", these guide a newly generated image and do not preserve input pixels; to modify an existing image, use type: "image_edit" with a source image. Reference and editing contract
The API guide documents validation constraints, including a portrait-only requirement for its explicit manga style preset.
A rejected request may therefore be a parameter mismatch rather than evidence that the model cannot create the requested scene. Validation rules
Keep named releases and API identifiers separate in project records. Luma announced the Uni-1.1 API on May 5, 2026, while the current Agents generation guide exposes uni-1 and uni-1-max as request values.
Use the identifier specified by your chosen endpoint. API announcement, Request documentation
For image review, treat reference matching as a requirement to test. Check object count, placement, colors, and fine details independently.
If a particular feature is essential, name it in the brief and inspect it in every output you consider delivering.
How do you use Luma Uni-1 in the app or Agents API?
For the browser workflow, sign in to the Luma application, open image generation, and select Uni-1 where offered.
Luma's product FAQ documents that route and describes adding detailed direction about camera, lighting, and style. Application instructions
Begin with a text-only version of your brief. Keep the subject, setting, viewpoint, and required relationships in separate sentences. This gives you a readable baseline before introducing references or more specialized output controls.
Next, supply a layout reference and explain its role: create a new room image with the same furniture arrangement and window position, using blue linen upholstery. Treat this as a reference-guided generation exercise; choose image editing if the deliverable requires modifying the source image itself. Reference contract
Save the reference, prompt, and result together. Compare furniture placement, window position, and upholstery against the brief before accepting a candidate.
For developers, the current Agents API documents the following request shape. Configure LUMA_AGENTS_API_KEY in your environment before running this example. Generation request documentation
curl --fail-with-body https://agents.lumalabs.ai/v1/generations \
-H "Authorization: Bearer $LUMA_AGENTS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "uni-1",
"type": "image",
"prompt": "A reading chair left of a floor lamp beside a tall window.",
"aspect_ratio": "3:2",
"output_format": "png"
}'
The initial response identifies a generation job. Poll GET /v1/generations/{id} until its state is completed or failed; on completion, the response's output entries contain download URLs.
The documentation provides the response fields and retrieval behavior. Response contract
Record the submitted model identifier and the returned job ID with your output. Download the accepted image into your project archive.
Treat the API response as evidence of which request ran, rather than assuming a marketing name identifies a fixed backend forever.
Once the basic request works, add reference inputs following the same guide's image_ref schema. Introduce one reference role at a time and keep a text-only baseline for comparison.
This suggested sequence helps isolate whether a reference improved the intended feature.
How does Luma Uni-1 compare with FLUX.2 Max?
| Model | Documented approach | Useful test |
|---|---|---|
| Luma Uni-1 | Unified understanding and generation with reference-based direction. Technical overview | A scene transformation with explicit relationships to preserve. |
| FLUX.2 [max] | Hosted generation and editing with multi-reference inputs and grounding search. BFL overview | The same reference brief with the same preservation checklist. |
See the sibling FLUX.2 Max guide for its service workflow. The ComfyUI complete guide provides broader context for organizing image-processing steps.
Evaluate both services against the deliverable you need. Keep input images and literal instructions constant, record any unsupported controls, and judge the result before considering cost or waiting time.
These are suggested comparison criteria, not a claimed benchmark outcome.
What are common Luma Uni-1 reference and API questions?
Can I download Luma Uni-1 weights?
Luma's Uni-1 product materials document hosted application and API access without an open-weight download. Use the Luma application or the documented hosted endpoint to generate images. Product documentation, API guide
How do I use multiple reference images with Luma Uni-1?
The Luma Agents API accepts up to nine image_ref entries for new image generation. State each reference's role in the prompt, and use type: "image_edit" with source when the task is to modify an existing image. Reference guide
Does a Luma Uni-1 image sequence mean video generation?
Luma's Uni-1 technical overview demonstrates storyboards and sequences of still images. Those examples describe image outputs; use a documented video workflow when your required deliverable is a video. Technical overview
Which Luma Uni-1 model name belongs in an API request?
The Luma Agents generation guide accepts uni-1 and uni-1-max in the model field. Luma's May 5, 2026 announcement uses the release name Uni-1.1, so follow the request values specified for the endpoint you call. API guide, API announcement
Sources
- Luma Uni-1 technical overview
- Luma Uni-1 product and usage FAQ
- Luma Agents image-generation API guide
- Luma Uni-1.1 API announcement
- Luma company announcement identifying Uni-1's launch year
- Black Forest Labs FLUX.2 overview
Top comments (0)