Scenario Luma Image
Overview
Uni-1, Luma Labs' image family on Scenario, folds generation and editing into one contract: every
member is both txt2img and img2img, and passing a source image is what flips the run into edit
mode, so where each image lands (source versus reference) decides more than prompt wording. These
are reasoning models that plan lighting and composition before rendering, so a run takes a minute or
two, not seconds. Discover members with search and treat model_schema_get as the contract: the
tiers share every field name and disagree on caps and price. Luma's video models are the
scenario-luma-video skill's domain.
Connection and the core loop: see the scenario skill in this repo; model-agnostic image work
(sizing families, masks, batch fields): the scenario-image skill. If a sibling skill named here is
missing from your available skills, load it with skill_get (Scenario MCP) or ask the user to
install it (https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas
and flag the gap.
Quick reference
Mode follows from the inputs (names from the live schema):
| Mode |
Inputs |
Behavior |
| Create |
prompt (+ imageRef) |
full scene from the prompt; aspectRatio honored, default 3:2 |
| Edit |
source + prompt (+ imageRef) |
prompt states the change; output keeps the source ratio unless aspectRatio is set |
imageRef is an array of guiding images and combines with source. Caps are per member: at
authoring time the Max hit took 9 references and the standard 8, with the source occupying one of
those slots when editing; both took a 6000 character prompt, nine aspectRatio values from 1:3 to
3:1, and outputFormat png (default) or jpeg. Each reference adds cost, so re-estimate with
dry_run after changing the count. webSearch (default false) has the model fetch real-world
visuals before generating: enable it when the prompt names a real place, product, or style the model
may not know. No seed, mask, pixel-size, or batch-count field exists: sizing is the ratio alone,
masked edits belong to other models, and identical re-runs cannot be pinned, so change one thing per
iteration. At authoring time the Max tier cost roughly two and a half times the standard for one 2K
image despite near-identical public arena ratings, so dry_run the same job on both before a batch.
References work by role
The model follows a reference reliably only when the prompt says what to take from it: character
likeness, style, composition, color palette, lighting, texture, or mood. Unlabeled references get
guessed at, and there is no adherence slider; influence rises with prompt specificity ("use the
first reference for the exact colorway and stitching"). Roles stack across references, one each.
Reusing one canonical reference across iterations is what holds a character steady.
Create prompts describe, edit prompts preserve
Create prompts read as one scene in natural prose: subject, setting, lighting, mood, style, and
always name the lighting, the single biggest quality lever. For text in the image, put the exact
string in quotes; rendered text is a family strength, and the Max tier's advertised edge is accurate
non-Latin scripts, not Latin text generally. Edit prompts are surgical: state the change first, then
pin what must not move ("Change X to Y. Keep Z exactly as it is."). One scene or one change per run.
Worked example: a travel poster with rendered title text
search with target="models", query="luma uni", public=true. Prefer the newest
non-deprecated hit, e.g. model_luma-uni-1-max (a live hit at authoring time: re-discover each
session).
model_schema_get with that id: reference cap, ratio list, and defaults before anything else.
upload_asset the palette reference (see the scenario skill) to get its asset id.
model_run with that model_id, dry_run=true, and the exact
parameters={"prompt": "A travel poster of Kyoto in autumn, a pagoda above red maples, warm golden hour light, flat-print texture. Use the reference for color palette and print grain. The title text \"KYOTO\" in bold serif across the top.", "imageRef": ["asset_a"], "aspectRatio": "2:3", "webSearch": true}
(a real place is named, so ground it).
- Repeat
model_run with wait=false, then jobs_wait with the returned job id, re-called with
pending_job_ids on timeout. The model reasons before rendering (median latency near 90 seconds
at authoring time), so a timeout is not a failure and never justifies a second model_run.
asset_display to check the title spelling and palette, then asset_download to save.
Common mistakes
- A style reference passed as
source: that flips the run into edit mode and the output hugs the
reference. References go in imageRef; source is only the image being changed.
- Unlabeled references: name each one's role in the prompt or the model guesses which to follow.
- An edit prompt with no preservation clause: whatever is not pinned is fair game.
- Expecting a square by default: create mode defaults to 3:2, so set
aspectRatio explicitly.
- Hunting for
seed, mask, width, or a batch count: none exist at authoring time; read
model_schema_get instead of assuming.
- Carrying the Max member's reference cap or price to the standard one: they share field names, not
numbers.
1---2name: scenario-luma-image3description: "Use when generating or editing images with Luma Uni-1 models on Scenario via MCP: text-to-image, prompt-based editing of an existing image, style or character reference images with named roles, web search grounding for real-world subjects, rendering exact title text into posters, aspect ratio control, or choosing between Uni-1 Max and Uni-1. Keywords: Luma Labs, Uni-1, Photon, txt2img, img2img, image edit, reference images, webSearch, reasoning image model."4license: MIT5metadata:6 source: https://skillmd.com/skills/scenario-labs/scenario-luma-image7---89# Scenario Luma Image1011## Overview1213Uni-1, Luma Labs' image family on Scenario, folds generation and editing into one contract: every14member is both `txt2img` and `img2img`, and passing a `source` image is what flips the run into edit15mode, so where each image lands (source versus reference) decides more than prompt wording. These16are reasoning models that plan lighting and composition before rendering, so a run takes a minute or17two, not seconds. Discover members with `search` and treat `model_schema_get` as the contract: the18tiers share every field name and disagree on caps and price. Luma's video models are the19`scenario-luma-video` skill's domain.2021Connection and the core loop: see the `scenario` skill in this repo; model-agnostic image work22(sizing families, masks, batch fields): the `scenario-image` skill. If a sibling skill named here is23missing from your available skills, load it with `skill_get` (Scenario MCP) or ask the user to24install it (https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas25and flag the gap.2627## Quick reference2829Mode follows from the inputs (names from the live schema):3031| Mode | Inputs | Behavior |32| ------ | ---------------------------------- | ----------------------------------------------------------------------------------- |33| Create | `prompt` (+ `imageRef`) | full scene from the prompt; `aspectRatio` honored, default 3:2 |34| Edit | `source` + `prompt` (+ `imageRef`) | prompt states the change; output keeps the source ratio unless `aspectRatio` is set |3536`imageRef` is an array of guiding images and combines with `source`. Caps are per member: at37authoring time the Max hit took 9 references and the standard 8, with the source occupying one of38those slots when editing; both took a 6000 character `prompt`, nine `aspectRatio` values from 1:3 to393:1, and `outputFormat` png (default) or jpeg. Each reference adds cost, so re-estimate with40`dry_run` after changing the count. `webSearch` (default false) has the model fetch real-world41visuals before generating: enable it when the prompt names a real place, product, or style the model42may not know. No seed, mask, pixel-size, or batch-count field exists: sizing is the ratio alone,43masked edits belong to other models, and identical re-runs cannot be pinned, so change one thing per44iteration. At authoring time the Max tier cost roughly two and a half times the standard for one 2K45image despite near-identical public arena ratings, so `dry_run` the same job on both before a batch.4647## References work by role4849The model follows a reference reliably only when the prompt says what to take from it: character50likeness, style, composition, color palette, lighting, texture, or mood. Unlabeled references get51guessed at, and there is no adherence slider; influence rises with prompt specificity ("use the52first reference for the exact colorway and stitching"). Roles stack across references, one each.53Reusing one canonical reference across iterations is what holds a character steady.5455## Create prompts describe, edit prompts preserve5657Create prompts read as one scene in natural prose: subject, setting, lighting, mood, style, and58always name the lighting, the single biggest quality lever. For text in the image, put the exact59string in quotes; rendered text is a family strength, and the Max tier's advertised edge is accurate60non-Latin scripts, not Latin text generally. Edit prompts are surgical: state the change first, then61pin what must not move ("Change X to Y. Keep Z exactly as it is."). One scene or one change per run.6263## Worked example: a travel poster with rendered title text64651. `search` with `target="models"`, `query="luma uni"`, `public=true`. Prefer the newest66 non-deprecated hit, e.g. `model_luma-uni-1-max` (a live hit at authoring time: re-discover each67 session).682. `model_schema_get` with that id: reference cap, ratio list, and defaults before anything else.693. `upload_asset` the palette reference (see the `scenario` skill) to get its asset id.704. `model_run` with that `model_id`, `dry_run=true`, and the exact71 `parameters={"prompt": "A travel poster of Kyoto in autumn, a pagoda above red maples, warm golden hour light, flat-print texture. Use the reference for color palette and print grain. The title text \"KYOTO\" in bold serif across the top.", "imageRef": ["asset_a"], "aspectRatio": "2:3", "webSearch": true}`72 (a real place is named, so ground it).735. Repeat `model_run` with `wait=false`, then `jobs_wait` with the returned job id, re-called with74 `pending_job_ids` on timeout. The model reasons before rendering (median latency near 90 seconds75 at authoring time), so a timeout is not a failure and never justifies a second `model_run`.766. `asset_display` to check the title spelling and palette, then `asset_download` to save.7778## Common mistakes7980- A style reference passed as `source`: that flips the run into edit mode and the output hugs the81 reference. References go in `imageRef`; `source` is only the image being changed.82- Unlabeled references: name each one's role in the prompt or the model guesses which to follow.83- An edit prompt with no preservation clause: whatever is not pinned is fair game.84- Expecting a square by default: create mode defaults to 3:2, so set `aspectRatio` explicitly.85- Hunting for `seed`, `mask`, `width`, or a batch count: none exist at authoring time; read86 `model_schema_get` instead of assuming.87- Carrying the Max member's reference cap or price to the standard one: they share field names, not88 numbers.