# Gpt Image

> Generate high-quality visual assets with GPT Image 2 / OpenAI image models for presentations, academic figures, diagrams, posters, UI mockups, and image-based slide backgrounds. Use when the user wants a PPT to look better, needs a visual style direction, slide background, method diagram, infographic, or image asset that can later be placed into a PPTX/HTML/PDF.

- Skill: `ricardo-vae/gpt-image` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add ricardo-vae/gpt-image`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ricardo-vae/gpt-image/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Web & Frontend
- Author: ricardo-vae (https://skillmd.com/u/ricardo-vae)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ricardo-vae/gpt-image

---


# GPT Image Skill

Use this skill when the task needs image-generation assistance, especially for:

- Research presentation visual redesign.
- PPT cover / section divider / background image generation.
- Academic method diagrams and visual abstracts.
- Infographics that explain experimental pipelines or results.
- Concept art or visual metaphors for research ideas.
- Image editing from a reference image.

This skill is best used together with `pptx`:

- `gpt-image` creates polished visual assets.
- `pptx` places those assets into editable PowerPoint slides and overlays editable titles, tables, metrics, and notes.

Avoid making the whole PPT a stack of flat screenshots unless the user explicitly wants that. For group meetings and thesis work, keep important text, tables, numbers, and captions as editable PowerPoint elements.

## Requirements

The usual execution path is one of:

```bash
gpt-image --help
uvx gpt-image --help
python ~/.codex/skills/gpt-image/scripts/generate.py --help
```

The API key normally comes from:

```bash
export OPENAI_API_KEY="..."
```

If the user uses a compatible relay / proxy endpoint, pass or export:

```bash
export OPENAI_BASE_URL="https://your-relay.example/v1"
```

This local machine is configured to auto-load:

```bash
~/.codex/gpt-image.env
```

That file contains the user's GPT Image API key, base URL, and default image model. Do not print the API key in chat. The current local default model is read from `GPT_IMAGE_MODEL`.

Do not ask the user to paste an API key into chat. Tell them where to set it locally.

## Recommended Workflow For PPT

1. Understand the presentation goal, audience, and slide count.
2. Decide which slides really need image assets.
3. Generate a style board first: one cover background and one content-page visual.
4. Ask for confirmation if the visual direction has major consequences.
5. Generate final assets in a project subfolder, e.g. `assets/gpt-image/`.
6. Use `pptx` to build or revise the `.pptx`.
7. Keep editable text over generated visuals.

For the user's current VLN/NavCoT work, good candidate image assets are:

- Cover background: indoor navigation map, panoramic rooms, path graph, soft academic style.
- Method framework: route pool -> visual grounding -> topology cleanup -> stop-memory.
- Bottleneck diagram: candidate oracle headroom vs confidence selection risk.
- Stop-memory diagram: visited trajectory, overrun stop, conservative back-1/back-2 correction.

## Prompting Rules

Good prompts should specify:

- Objective: what the image is for.
- Format: PPT background, diagram, infographic, icon set, etc.
- Style: academic, clean, modern, high-end conference talk, not cartoonish.
- Content: the visual objects that must appear.
- Constraints: leave empty space for editable text, no dense tiny labels if text will be added in PPT.
- Color direction: match the user's template when relevant.
- Output size: 16:9 if it is for slides.

For PPT backgrounds, prefer prompts that request "no embedded text" or "minimal abstract labels only"; add final text in PowerPoint.

## Example Commands

If `gpt-image` CLI is installed:

```bash
gpt-image generate \
  --model gpt-image-2 \
  --size 1536x864 \
  --output assets/gpt-image/cover.png \
  "A polished academic presentation cover background about vision-language navigation..."
```

If only `uvx` is available:

```bash
uvx gpt-image generate \
  --model gpt-image-2 \
  --size 1536x864 \
  --output assets/gpt-image/cover.png \
  "A polished academic presentation cover background about vision-language navigation..."
```

If CLI options differ, first run:

```bash
uvx gpt-image --help
```

Fallback helper in this skill:

```bash
python ~/.codex/skills/gpt-image/scripts/generate.py \
  --prompt "A clean academic 16:9 slide background..." \
  --output assets/gpt-image/cover.png \
  --model gpt-image-2 \
  --size 1536x864
```

## Design Guidance

For academic PPT:

- Use generated images as atmosphere, visual metaphors, or diagrams, not as all-text slides.
- Prefer a small set of consistent motifs across the deck.
- Use slide-safe negative space for titles and metric callouts.
- For diagrams, keep labels short or omit labels and overlay editable labels later.
- Avoid dense AI-generated pseudo-text, fake logos, fake paper screenshots, and unreadable micro labels.

For the user's existing Zhejiang Sci-Tech / IPCE style:

- Suggested palette: white, warm yellow, near-black, muted blue, soft gray.
- Keep university/lab logos as original assets, not regenerated.
- Use generated assets behind content or inside framed panels.

## When To Combine With Other Skills

- Use `pptx` after image generation to create editable `.pptx`.
- Use `paper-source-finder` / `deep-research` before image generation when content accuracy matters.
- Use `pandoc` only for document conversion, not for visual slide design.