GPT Image 2.5 · skill pubblica Open Skill
GPT Image 2.5
You can leverage this skill to drive GPT Image 2.5, harnessing its powerful capabilities to assist you in image generation tasks. This integrated workflow allows you to seamlessly trigger the image model through the skill framework, eliminating cumbersome manual operations and streamlining your creative process. Whether you need to craft detailed visual prompts, adjust artistic styles, control composition parameters, or iterate on generated visuals, GPT Image 2.5 can be activated directly by this skill to produce tailored images that match your requirements. It serves as a convenient bridge between your instructions and the image generation engine, enabling flexible real-time adjustments to refine outputs. With this setup, you no longer need to switch between separate tools; you can submit your creative requests and receive high-quality generated images efficiently, greatly boosting productivity for design, concept visualization, and other visual creation work.
Other
154 21 53
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-09-16/ak52p38r.jpg
# GPT Image 2.5
Generate or edit images with **GPT Image only**. No structured project state. Speak to the user in the user's language.
**Channel lock:**
| Mode | When | Primary `model` | Fallback `model` (only after two primary failures) | `input_urls` |
|---|---|---|---|---|
| **t2i** | New picture from text; no source image for this call | `t2i.hd.gpt` | `t2i.hd.lite` | omit |
| **i2i** | User upload, reference image, or edit/keep identity from the last result | `i2i.hd.gpt` | `i2i.hd.lite` | required |
Always start with the **primary** GPT channel. Use the matching **lite** channel only as fallback after two consecutive primary failures for this same request (see **Generation failures**). When falling back, tell the user: `I'll generate the image with nanobanana for you.` Never use any other channel (`t2i`, `t2i.hd`, `i2i`, `i2i.hd`, …). Never omit `model`. Never pass `input_urls` with t2i. Never omit `input_urls` with i2i. Do not ask the user which model to use.
Prompt craft lives in `references/prompt-templates.md`. **Do not read it up front.** Read it with `read_skill_file` only when you are about to write a `generate_image` prompt.
## When to use
- User wants a new image from a text description
- User uploaded image(s) and wants to edit, restyle, combine, or generate from them
- User uploaded image(s) but has not said what they want yet → use this skill, **understand intent first, then wait**
- Follow-up tweaks on the last image from this skill
- User wants to change the ratio of the last image, or upscale / raise resolution / make it sharper
- User named GPT / GPT Image / gpt-image-2.5
Do not use for video, voiceover, music, scripts, research, or small talk. Do not generate if they only ask what an image is / whether it looks good.
## Size and ratio
**Do not ask the user for ratio. Do not choose a ratio yourself.** On a normal generate, omit `orientation` and `aspect_ratio`. Do not put ratio or pixels in the prompt. Never pass pixel width/height. Never pass `resolution` to `generate_image` to make a picture larger or sharper.
Ratio is only applied when the user **later** asks to adjust it (taps `Adjust ratio` or names a ratio after there is already an image). Then pass both fields from this table — still do not ask before the first generate:
| User choice | `orientation` | `aspect_ratio` |
|---|---|---|
| Landscape 16:9 | `landscape` | `'16:9'` |
| Portrait 9:16 | `portrait` | `'9:16'` |
| Landscape 4:3 | `landscape` | `'4:3'` |
| Portrait 3:4 | `portrait` | `'3:4'` |
| Square 1:1 | `square` | `'1:1'` |
Only these five pairs. Other ratios → say only these five are supported.
## Upscale
If the user wants higher resolution, sharper output, upscale, enlarge, 2K/4K, or taps `Upscale`: call `upscale_image`. **Do not** re-generate with a higher `resolution`.
- `input_url`: last successful image from this skill (or the upload they named). Never invent URLs.
- `scale`: `2` by default; `4` if they asked for 4x / much larger / 4K
- `name`: short title, e.g. `upscaled`
- `entity_type='edited_images'`
- No image yet → generate first (or ask for an upload); do not upscale a missing URL
## After-success guidance
After every successful `generate_image` or `upscale_image`, in the **same turn** as the final reply, call `set_guidance` with **these two first**, in this order:
```
set_guidance(suggested_questions=[
{"id": "adjust-ratio", "text": "Adjust ratio"},
{"id": "upscale", "text": "Upscale"}
])
```
Do not call other tools in that turn. Extra buttons may follow these two; do not put anything before them.
- Tap / say **Adjust ratio**: do not generate yet. Same turn, `set_guidance` with **only** the five ratio buttons, then wait. After they pick (or name a pair), regenerate **i2i** from the last successful image with that `orientation` + `aspect_ratio` (primary GPT channel; lite fallback still applies). Then success guidance again (`Adjust ratio` then `Upscale`).
- Tap / say **Upscale** or any raise-resolutio