The Master Guide / 06 Anchors and stills · Still prompts, storyboard frames and diagrams
GPT Image's two jobs and its version rule
GPT Image keeps only the two jobs the other stills models don't do: it generates with a transparent background, and it draws exact marks and short labels (decided).
GPT Image keeps only the two jobs the other stills models don't do: it generates with a transparent background, and it draws exact marks and short labels (decided).
The two jobs (decided): transparent cut-outs, and graphics with text. Every staging map and position image, labelled or not, counts as a graphic with text, so one tool makes every diagram (decided).
The version rule (decided): a new project uses the latest GPT Image version on the platform, as with Suno. A project stays on the version its first approved GPT Image asset was made on, so its assets never change version mid-film. Record the version in your plan. On 5 Nov 2026 the latest is GPT Image 2.6.
Setting | Choose |
Model | Pick it by name on every job, at the project's version. On 2.6: the standard model, Sunburst, for precision (diagrams, maps, photoreal cut-outs), and its Flare model for stylised cut-outs. On 2.5: one model with a variant setting, Sunburst for precision and Flare for stylised [official guidance] |
Quality | Always high; the highest setting for dense text. Never the default, which silently returns low quality at 1K |
Resolution | 2K (the default is 1K) |
Background | Transparent, for cut-outs only |
Output | PNG, one image per job. Not GIF or SVG: moving graphics are built in code (see Graphics, captions, versions and delivery → Graphics built in code) |
Aspect | Always set: the shot's ratio for a map drawn from its camera; 16:9 for top-down maps and maps of a place; the use's ratio for cut-outs |
References | Numbered, each with its role, in the order you attach them. Up to 16 on 2.5; 2.6 states no limit, so keep to 4–6 |
Planning | On 2.6 you can't set how much it plans, and it plans heavily, which can add changes you didn't ask for in an edit: check every edit at full size |
Cost | 2.5: about 2.75 Higgsfield credits (cr) at 2K high, 4.25 cr at 4K. 2.6: no price recorded yet. Before its first use, read the credit cost the Generate button shows for one job at the project's settings (model, quality, resolution) before you generate, and write it with the date in your price ledger (see Planning, budgets and takes → Costs and the estimate) |
Prompt style [official guidance]: the result first, then short labelled sections.
The first line: what the image is, its use, the view and the aspect ("A staging diagram, front view from the shot's camera, 16:9").
INPUTS: each image numbered, with what it owns and what to ignore.
Labelled sections, only those needed: CONTENT, LEGEND, STYLE, TEXT, LOCKS.
About 80–150 words, standing alone, with no "as above" and no quality tags.
The text lock: no text, letters or logos anywhere, unless the TEXT section lists the exact labels.
Attach the approved source of a cut-out every time (the client's photo, the approved stylised asset), never a near-miss or an earlier output. No real person's name in any prompt, and never anything made from a storyboard frame.
Edits follow the edit budget (see Planning, budgets and takes → Fixing a failed take): pass the last output in, name the one change, list what stays, and restate the constraints if anything drifts [official guidance]. On 2.6, compare the whole image with the one before, not only the changed area. No pixel fixes (decided).
Symptom | Likely cause | One change for the next attempt |
A soft, low-detail result | Settings left at the default (low, 1K) | Set the model, high quality and 2K explicitly |
A job ran on the wrong version | The model left unnamed, and the default moved | Name the project's recorded model every time |
A diagram draws labels or arrows | No text lock | Rule out labels, numbers, letters and arrows in LOCKS |
A wrong or extra label on a map of a place | Labels not listed exactly | A TEXT section with the exact labels and nothing else |
Figures in the wrong colours | Colours not written | Name each figure's colour in CONTENT, and in LEGEND where the plan uses one |
Stray letters or a fake logo | The model fills space with text | The text lock in LOCKS; any real text comes from the client's artwork |
A shadow or halo on a cut-out | Transparency not set, or a shadow implied | Set the transparent background; rule out shadow and reflection |
A stylised cut-out in another style | The source asset not attached | The approved stylised asset as Image 1, owning design and style |
An edit changed more than asked (2.6) | The model's planning | Name the one change and list everything that stays; regenerate if it happens again |
Sources
OpenAI's image-prompting guide (the result first, labelled sections, numbered inputs, one change per edit); Higgsfield's GPT Image 2.5 page (the Sunburst and Flare variants).