2.9 KiB
GPT Image 2.5: generation
Read for new photos, illustrations or product scenes. For existing-image changes, use editing instead. Source check: 2026-09-09; provenance and validation boundary.
Official examples to consult selectively
Open only the matching example, not the whole guide. Camera specifications are appearance cues, not exact physical simulation. The examples do not establish which variant will win on the user's brief.
Local workflow using the existing gallery and craft
These are repo application notes, not locally measured 2.5 performance claims. Start from the user's brief; consult a gallery case only for missing inspiration. Do not append every possible quality adjective.
| Deliverable | Resolve before generation | Inspect in the actual output |
|---|---|---|
| Natural photograph | Subject/action, framing, light direction, texture and retouching limits | Skin, hands, contact shadows, reflections and background perspective |
| Product scene | Product features, placement, materials, scene palette, required label copy | Geometry, label spelling, reflections and contact with the surface |
| Illustration | Medium, shape language, composition, palette and intended use | Consistent medium, silhouette readability, unwanted photorealism |
| Logo concept | Approved name, symbol, use size and background | Spelling and small-size legibility; a raster concept is not editable vector artwork |
If a capture/material question remains, consult the relevant section (8, 10 or 12) of craft.md, not the full file. For type-led deliverables, switch to layout and text. If a reusable portrait starting point is needed, see the Natural editorial photograph section of community templates, retaining its source label.
Before calling the CLI
- Resolve factual details from the brief; do not inherit dates or claims from a historical gallery example.
- Use the confirmed model and settings. Description words such as “4K” do not set API dimensions; check models.md.
- Do not add a reference-identity promise to text-only generation. Attach the relevant input and follow the editing slice when identity must be preserved.
- State the failed visual requirement before proposing a paid refinement; request permission rather than automatically increasing quality or switching models.