GPT2.5 Image Editing · публичный навык Open Skill

GPT2.5 Image Editing

GPT2.5 Image Editing — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.

@eggs

Other

4 1 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/xa8gwh5j.jpg

# GPT2.5 Image Editing

Turn an uploaded picture plus a text instruction into a finished image rendered in one single "gpt2.5 image editing" visual treatment. There is no structured project state: every turn needs the reference image (uploaded by the user or already on the canvas) and the user's instruction.

## When to use

- The user uploads an image and asks for a result image: "edit this", "rework this photo", "make it look like gpt2.5 image editing", "change the background / material / lighting".
- The user uploads an image and adds text describing what should appear in or what should change about the picture.
- The user gives only text but explicitly asks for the gpt2.5 image editing style.

## Steps

1. Collect the inputs for this turn:
   - Reference image(s): the user's upload or an asset already on the canvas (use its `asset_id` or `s3://` uri).
   - Instruction text: what the user wants changed, added, kept, or emphasized.
   If there is no reference image at all, ask for it once. If the user insists on text only, continue with the text-to-image path in step 4.

2. Before any generation call, read the style and prompt conventions:
   `read_skill_file(name='gpt25-image-editing', path='references/style-and-prompt.md')`
   Use it for color, material, lighting, layout, keep/avoid lists, and the prompt template. Do not generate before reading it.

3. Write the final prompt in English, following that reference file. Always build it as explicit blocks: STYLE, SUBJECT/content (derived from what is actually visible in the uploaded image plus the user's instruction), LAYOUT, MUST KEEP, MUST AVOID. Requirements the user stated as non-negotiable go into MUST KEEP verbatim; anything the reference file forbids goes into MUST AVOID. If the picture must contain literal on-image text, keep that text in its original language and quote it exactly.

4. Call `generate_image`:
   - Reference image present (the normal case): `model='i2i.hd.gpt'`, `input_urls=[<reference image uri or asset_id>]` — the first entry is the base image being edited — plus the prompt built in step 3.
   - Text only, no image available: `model='t2i.hd.gpt'`, prompt = STYLE block + described subject/layout blocks.
   - `orientation`: match the reference image's aspect ratio (`portrait` / `landscape` / `square`) unless the user asks for a different one; for text-only requests default to the ratio that suits the described subject.
   - `resolution`: pass `'1080P'` when fine detail, material texture, or on-image text must survive the edit; otherwise omit it and use the platform default.
   - `name`: a short readable title, e.g. "GPT2.5 edit - <subject>". `entity_type`: `edited_images`.
   - Never pass a physical model name; the channel strings above are the contract.

5. Reply in one or two plain sentences naming what was changed and what was preserved. The result is already on the canvas, so do not paste URLs. If a further step is obvious (another edit pass, a different aspect ratio, a variant), call `set_guidance` in the same turn as the final text.

## Notes

- Preserve what must survive: the identity and recognizable structure of the original subject, the pose and framing, and any detail the user explicitly asked to keep. Never silently remove or replace those.
- Do not invent content that is neither visible in the reference image nor stated by the user or the app description: no extra people, props, logos, text, or scene elements.
- If the user's instruction conflicts with the default keep rules, the user's explicit instruction wins for that element only; the rest of the conventions still apply.
- Multiple edits in a row: use the previous output as the new reference image for the next `generate_image` call.
- The `.hd` channels used here do not support NSFW or adult content. Politely decline such requests and do not call the tool.
- If `generate_image` returns `success=false`, tell the user the actual failure plainly; never claim success or fabricate an image li