Scenario Gpt Image

Scenario GPT Image

Overview

Every member of GPT Image, OpenAI's image family on Scenario, both generates and edits through one prompt, with referenceImages carrying the image to edit and an optional alpha mask for inpainting. Four members were live at authoring time: GPT Image 2.5 Flare (fast, everyday), GPT Image 2.5 Sunburst (precision: dense copy, diagrams, edits that must hold geometry), GPT Image 2, and GPT Image 1.5, the one member with a fidelity dial and ratio sizing. Discover members with search and treat model_schema_get as the contract: the family agrees on prompt and references and splits on sizing, quality tiers, and fidelity.

Connection and the core loop: see the scenario skill in this repo; model-agnostic image work: the scenario-image skill. 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

Shared shape at authoring time: prompt (required, cap 32000 characters), referenceImages (an array even for one, up to 10), numOutputs (1 to 10 variations of one prompt), quality, and background (auto and opaque everywhere, transparent on every member but GPT Image 2). The splits:

Routing: Flare and Sunburst share one contract and one price list, so the choice is speed against quality. OpenAI describes Flare as the small model, built for speed with image quality comparable to GPT Image 2, and Sunburst as the base model, with higher image quality than GPT Image 2: Flare for everyday work, Sunburst when accuracy beats speed (packaging copy, dense numbers, exploded diagrams, geometry-preserving edits). GPT Image 2's contract is the 2.5 contract minus the two top tiers, and it prices its high at their max, so new work starts on 2.5. GPT Image 1.5 earns its place for inputFidelity: "low", which reworks a reference instead of reproducing it (only the first five references keep the higher fidelity), and for 3:2 or 2:3 output without pixel math. Transparent cutouts route to a 2.5 member: GPT Image 2 rejects the value, and 1.5 is the fidelity and ratio member rather than the cutout one.

Quality is the price dial

quality moves the price more than size does, and the same word does not price the same across members. At authoring time a 1024 square priced 2, 3, 12, 20 and 45 CU across low, medium, high, xhigh and max on both 2.5 members; GPT Image 2 priced low at 2, medium at 12 and high at 45; auto resolved to 12 on all three; a 3840 by 2160 max run on Flare came to 69 CU. dry_run=true the member you will run and read creativeUnitsCost; referenceImages, numOutputs, quality, and sizing all carry cost_impact, so re-estimate after touching any. xhigh or max on GPT Image 2 is a 400 ("Input quality must be one of the following values") even on a dry run. Tier choice: low for drafts, medium for everyday finals, high when small text or legends must stay legible, and xhigh or max only when the tier below left a specific requirement unmet within your latency budget: a higher tier does not guarantee a better result for every prompt.

Sizing on the pixel members follows OpenAI's published grid: each axis a multiple of 16, long edge at most 3840, aspect between 1:3 and 3:1, total pixels between 655,360 and 8,294,400, above 2560 by 1440 experimental. The schema enforces only the 16-step range and dry_run prices any request, so hold to the grid yourself: an off-grid request still succeeds but returns a different size with no error. At authoring time 3840 by 960 and 3840 by 480 both came back 3840 by 1280 (clamped to 3:1), 960 by 3840 came back 1280 by 3840, 1000 by 1000 came back 992 by 992, and 3840 by 3840 came back 2880 by 2880 on Sunburst while Flare refused it at submission ("Error while running prediction"). Common picks: 1024 square (fastest), 1536 by 1024, 1024 by 1536, 3840 by 2160 for 4K, 2560 by 1088 for a true 21:9.

An edit source outside 1:3 to 3:1 comes back padded, not cropped: a 1326 by 313 (4.24:1) source returned about 2132 by 738 on both 2.5 members with white bands above and below covering about a third of the frame, and typing the source size returned 1408 by 480 padded the same way. Read width and height from asset_get after every run. When the edited file must keep its exact canvas, crop the bands off the result or route the edit to a member that keeps the source size (see scenario-ideogram).

Masks and transparency

The mask is a PNG the size of the first reference image, read through its alpha channel only: transparent pixels are edited, opaque pixels preserved, RGB white or black ignored without alpha. Export with transparency, never flattened, and convert a white-on-black mask's white region to transparent. OpenAI treats the mask as guidance rather than a hard edge: when a region must stay pixel-identical, composite the approved edit back over the original.

background: "transparent" returns real alpha on 2.5 Flare and Sunburst at every tier; GPT Image 2 rejects it. Two authoring-time observations on the 2.5 members: alpha peaks at 254, never 255, so a check for fully opaque pixels reads the whole subject as semi-transparent; and the color data under transparent pixels carries a halo, so any surface that drops alpha (JPEG flattening, a thumbnail without transparency, a viewer compositing on black) shows the cutout glowing. On white or a checkerboard it is clean. Prompt the cutout as a clean cutout with a crisp silhouette, fine edges and label text preserved, no halos, no restyling, then inspect the alpha channel at hair, glass, shadows, and object edges. Keep PNG or WebP and the alpha channel through every hop, previews and downloads included.