MetaSkill/skills/le-gpt-image/references/gallery/openai-image-2.5-editing.md

GPT Image 2.5: references and preservation

Read for reference edits, translation, inpainting and transparent cutouts. Source check: 2026-09-09; provenance and validation boundary.

Official examples to consult selectively

• Translation with layout preservation
• Identity and clothing
• Combining references
• Transparent product cutouts
• Character consistency across turns

Improved preservation is not a pixel lock. Repeated edits can drift; exact unchanged regions require compositing rather than relying on prompts alone. Official migration caution

Local workflow using the existing CLI

1. Inspect the actual inputs. Assign each a role: base scene, identity, garment, style or inserted object. Keep this order consistent with repeated -i flags.
2. Separate the requested change from invariants; do not turn a small edit into a restyle. Consult craft.md section 13 only if this needs clarification.
3. Keep the selected model explicit. Use models.md for API settings; do not infer that 2.5 requires --input-fidelity high.
4. Compare the changed region and the rest of the image. For another authorized edit, pass the actual approved output through -i; the CLI does not retain a conversation or attach an image merely because its filename is in the prompt.

Task	Inspect before accepting
Translation	Approved replacement copy, untranslated remnants, original layout and logos
Try-on / identity edit	Face, expression, pose, garment details and plausible body geometry
Multi-reference composition	Correct source for each inserted element, scale, perspective and lighting
Cutout	Product outline/labels, edge halos, hair/glass/shadows and real alpha transparency
Multi-step edit	Current requested change plus accumulated drift from the original invariants

Masks and transparent output

A mask is edit guidance, not a guaranteed hard pixel boundary. With multiple inputs it applies to the first image. Match mask dimensions to that image and check its alpha channel; use -i plus -m. See the official mask requirements for input constraints.

For cutouts, use --background transparent --format png (or WebP), not JPEG. Inspect the decoded alpha channel; a painted checkerboard is not transparency. The prompt's preservation list alone does not establish that anything was preserved. If the job requires exact untouched pixels, explain the need for a separate compositing step; do not claim the generative edit supplies that guarantee.

For a reusable change-only prompt, read Single-change preservation edit in community templates. API access and local output validation remain governed by models.md.