Scenario Ideogram Image

Overview

Ideogram's image family on Scenario is a set of single-purpose members, not one model with modes: 4.5 for generation where in-image text must read and for reference-guided edits, 4.5 Precise Edit for changing one thing and leaving the rest alone, V3 Generate Transparent for native alpha output, V3 Layerize Text for turning a flat graphic into editable text layers, Character for one identity held across scenes, and Remove Background for cutouts. V4 still answers searches but was tagged deprecated in favor of 4.5 at authoring time. The members agree on almost nothing mechanically: the same concept changes name, casing, and allowed values between them, so discover each with search and treat its model_schema_get as the contract.

Connection and the core loop: see the scenario skill in this repo; model-agnostic image work (sizing families, reference cardinality, masks): the scenario-image skill; running a bake-off against another family: scenario-model-comparison. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Pick the member by the job (names from the live schemas, caps at authoring time):

Member	Job	Inputs that matter
4.5	Text-heavy generation, ref edits	prompt, referenceImages (up to 5, the first is the source), mask, aspectRatio (auto, source, 25 ratios from 3:1 to 1:3), resolution (1K, 2K), quality, magicPrompt, numOutputs (up to 8)
4.5 Precise Edit	Change one thing, keep the rest	image (output keeps its size), mask, referenceImages (up to 4 guides), quality (default medium), numOutputs (up to 8)
V3 Generate Transparent	Native alpha output	prompt, negativePrompt, aspectRatio (15 ratios), renderingSpeed (adds FLASH), expandPrompt, numOutputs (up to 8), seed
V3 Layerize Text	Flat graphic to layers	image, optional prompt, fontName* or fontFile* per tier (H1, H2, Body, Small), seed
Character	Same character, new scenes	prompt, characterReferenceImage, styleType (Auto, Fiction, Realistic), aspectRatio or resolution, image plus mask, renderingSpeed (Default, Turbo, Quality), seed
Remove Background	Cutout to transparent PNG	image (10MB cap)

Nothing transfers between members. Masks flip between them: on 4.5 and Precise Edit white marks the region to edit, on Character black repaints and white stays. Only V3 Transparent takes a negativePrompt. renderingSpeed is uppercase on V3 Transparent and Title case on Character, whose middle tier is Default.

4.5: tiers, sizes, and what they cost

quality is the price dial on both 4.5 members (very_low, low, medium, high; very_low needs an input image, which Precise Edit's image satisfies and 4.5 takes as a referenceImages entry). At authoring time dry_run quoted 7.75, 13.75, and 41.75 CU for low, medium, and high, Precise Edit's very_low 3.75, and the same figure at 1K and 2K and for a 0.4 MP, 4.3 MP, or 8.3 MP source; billed jobs came in 1.75 CU under each estimate. 4.5 defaults to high, its top and most expensive tier; Precise Edit defaults to medium, and its high picks the best of several candidates at three times the price. Tier names do not line up across families: Ideogram's high is its ceiling, while GPT Image 2.5's high is its third tier of five and priced near Ideogram's medium. Compare families at a matched price, never a matched tier name, and dry_run the exact payload.

2K output observed at authoring time: 2048 by 2048 (1:1), 2560 by 1440 (16:9), 2880 by 1440 (2:1), 2944 by 1152 (23:9), 3072 by 1024 (3:1), so the long edge tops out near 3072 and there is no 4K. 1K is about 1 MP (16:9 returned 1280 by 720) and exists only for 1:1, 4:5, 3:4, 2:3, 5:8, 9:16, 1:2 and their landscape counterparts; 3:1 at 1K passes dry_run, then fails at run time with a hint naming the fix. A ratio outside the list (4:1) is a 400 listing the allowed values. A mask requires aspectRatio: "auto".

Edits that keep the canvas

For an edit, aspectRatio: "source" on 4.5 and every Precise Edit run return the source's exact pixel size, even one outside the ratio list: a 1326 by 313 (4.24:1) turnaround sheet came back 1326 by 313 from both, while auto rebuilt it as a full 3072 by 1024 scene. Use this when the edited file must drop back into a layout, a sprite sheet, or a page template unchanged. A localized change such as a headline swap needs no mask on Precise Edit: maskless at the medium default left the rest of the scene untouched. The GPT Image family pads such a source with white bands instead (see scenario-gpt-image).

In a single-sample bake-off at authoring time against GPT Image 2.5 on the same sources, the 4.5 members disturbed two to four times fewer pixels outside the edited region on a sign-text swap and an add-and-remove-objects edit, and read closer to a requested painterly medium (a watercolor restyle, a gouache landscape; hands-on use found the same for pencil sketches). GPT Image 2.5 led on photoreal product shots, on following a character brief, and on material and color changes: asked for burnished gold armor, both 4.5 members returned an olive brass, and Precise Edit also recolored the sword the prompt said to keep. Route by the job, and say "keep everything else exactly the same" plus the list of what must survive in every edit prompt.

Exact text wants expansion off

The generators rewrite the prompt before generating by default, which helps a short exploratory prompt and hurts the family's specialty: the rewrite can paraphrase the exact copy that must render. When the image carries wording, disable it (magicPrompt: "off" on 4.5, expandPrompt: false on V3 Transparent), quote each piece of copy, and give it a place and a style ('the headline reads "GRAND OPENING" in bold condensed capitals across the top'). magicPrompt applies to text-to-image only: 4.5 turns every edit into structured instructions regardless.

Two routes to transparency

Native: V3 Generate Transparent writes the alpha channel directly, so icons, stickers, and UI elements arrive compositing-ready. Cutout: generate on 4.5, then run Remove Background on the result; it reconstructs edge pixels with partial transparency rather than segmenting, so hair, fur, and glass survive. Go native when the asset is designed as an isolated element; go cutout when typography or overall quality leads. 4.5 has no background field, so prompting "transparent background" there yields an opaque image.

Worked example: a localizable poster

1. search with target="models", query="ideogram", public=true. Members return as separate hits; match by name, e.g. model_ideogram-v4-5 for typography (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: field names, allowed values, and defaults before anything else.
3. model_run with that model_id, dry_run=true, and parameters={"prompt": "Retro travel poster, warm dusk palette. The headline reads \"KYOTO IN BLOOM\" in bold serif across the top; caption \"April 2027\" bottom right.", "aspectRatio": "9:16", "resolution": "2K", "quality": "high", "magicPrompt": "off", "numOutputs": 2} for the cost estimate.
4. Repeat model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
5. asset_display the outputs and pick one.
6. To make the copy editable, discover the Layerize member the same way (model_ideogram-v3-layerize-text at authoring time), read its schema, and model_run with parameters={"image": "<poster asset id>"}: a generated asset's id feeds a file input directly, no re-upload. The output is a text-erased base plus text blocks with role, position, and content.
7. jobs_wait, then asset_display and asset_download.

Common mistakes

• Carrying one member's block to the next: magicPrompt and resolution are 4.5's; V3 Transparent wants aspectRatio and expandPrompt; Precise Edit takes image, not referenceImages, for the source.
• Leaving 4.5 on its high default for drafts: low or medium first, high for the final.
• Leaving expansion on with exact copy in the prompt: the rewrite can change the words before the model sees them.
• Editing with aspectRatio: "auto" when the canvas must not change: use source, or Precise Edit.
• Prompting a transparent background on 4.5: generate on V3 Transparent or chain Remove Background instead.
• Inpainting on Character with image but no mask: they go together, characterReferenceImage stays required, and sizing fields are ignored while inpainting.
• Setting fontNameH1 and fontFileH1 together on Layerize: a tier's font comes from a font name or a font file, never both. upload_asset has no font kind, so with only a local font file use the font name route and flag the gap.