GPT Image 2.5 Remake · публичный навык Open Skill

GPT Image 2.5 Remake

GPT Image 2.5 Remake — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.

@eggs

Custom

1 2 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/ewdah8cg.jpg

# GPT Image 2.5 Remake

Turn a user-uploaded reference image into a remade image. The remake keeps the identity, structure and layout of the original while rewrapping it in one single, consistent visual treatment (color, material, light). This is an image-to-image task: the uploaded picture is the source of truth, never a text-only prompt.

## When to use

- The user uploads a picture and asks to remake / rework / redraw / recreate / re-render it.
- The user asks to "remake this in GPT Image 2.5", "give me a new version of this photo", "keep the same subject, new look".
- The user wants to redo only one aspect (color, light, background) of a remade image already produced in this conversation.
- Not for: video, audio, or pure text-to-image requests with no reference image (see the fallback in step 4).

## Inputs and outputs

- Input: one or more reference images supplied by the user (an uploaded asset, a canvas asset, or an `s3://` uri / `asset_id` already present in the conversation).
- Output: one remade image, added to the canvas. The original image is never modified — every call produces a new asset.

## Steps

1. Identify the reference image. If the user has not provided any image and only described one, ask them to upload it before generating. If several images are given, put the main subject image first in `input_urls`.

2. Build the remake prompt in English from the template in "Prompt template" below. Fill in the subject sentence from the reference, keep the style block fixed (it is what makes the remake look like one coherent piece), and carry over every element the user asked to preserve.

3. Call `generate_image` with these conventions:

   - `prompt`: the filled English template (step 2)
   - `input_urls`: `[<reference image asset_id, preset key or s3:// uri>]` — the uploaded reference; more than one entry only if the user supplied several
   - `model`: `i2i.hd.gpt` (image-to-image remake channel; never write a physical model name)
   - `orientation`: match the reference — `portrait` for a vertical source, `landscape` for a horizontal one, `square` for a square source. Only change it if the user explicitly asks for a different frame shape.
   - `resolution`: omit it by default (platform default 720P). Pass `1080P` only when the user asks for higher clarity or the source is a detailed photo that clearly needs it.
   - `name`: a short canvas title, e.g. `remake-portrait`, `remake-product-card`
   - `entity_type`: `image_remakes`

4. Text-to-image fallback: if the user explicitly asks for a remake-style image with no reference image at all, generate it with `model='t2i.hd.gpt'`, `prompt` written from the same style template, and no `input_urls`. Do not use the fallback to avoid asking for an upload when the user clearly meant to give one.

5. Reply in the user's language with one or two sentences summarising what was remade (kept composition, unified color/material/light). Do not paste long URLs. If a follow-up would help (another look, a different frame shape), call `set_guidance` in the same turn.

## Prompt template

Fill the fields in `<>`; keep the fixed style block as is.

```
Remake the reference image as one unified piece in the GPT Image 2.5 remake look.

Subject: <one sentence describing the exact subject of the uploaded image — who/what it is, its pose or state, and its key objects>.

Style (apply to the whole frame, no exceptions): one single colour scheme — <base palette, e.g. clean neutral base with one saturated accent colour taken from the subject>; one single material treatment — <e.g. smooth matte surfaces with soft specular highlights, no mixed textures>; one single lighting setup — <e.g. one soft directional key light from the upper left with gentle falloff and clean ambient fill>. Colour, material and light must stay consistent edge to edge; the entire image serves this one theme. No mixed styles, no collage, no extra frames, no borders, no text overlays.

Layout: keep the original composition — s