# Image Gen

> Use when the user asks to generate, edit, use reference images, or batch-create raster images with any gateway image model. Uses YOUTU_API_KEY and the bundled image-jobs CLI. Do not use for SVG, pure CSS/HTML graphics, or image analysis only.

- Skill: `inspirepan/image-gen` (Agent Skill, multi-file: 9 files)
- Install (CLI): `npx skillmds@latest add inspirepan/image-gen`
- Raw SKILL.md: https://api.skillmd.com/api/skills/inspirepan/image-gen/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Web & Frontend
- License: Apache License 2.0
- Author: inspirepan (https://skillmd.com/u/inspirepan)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/inspirepan/image-gen

---


# Image Generation

Use the model-agnostic Youtu image-jobs gateway through `scripts/image_gen.py`.
It defaults to `gpt-image-2`, but supports any model exposed by the gateway.
Live requests require `YOUTU_API_KEY`; `YOUTU_BASE_URL` optionally replaces
`https://api.youtu.uk`. Run commands with `uv run`.

## When to use

- Create raster artwork, product images, covers, illustrations, or UI imagery.
- Edit an existing image, make a masked change, or combine local or HTTPS-hosted images.
- Make variants or a JSONL batch using a gateway image model.

## When not to use

- SVG/vector work, diagrams, simple CSS/HTML visuals, or native editable assets.
- Requests that only analyze, describe, or classify an image.

## Workflow

1. Decide whether the goal is **generate** (new asset or reference-guided asset) or
   **edit** (preserve an existing image while changing it). Use batch for
   distinct prompts. Use `--n` for variants only when the selected model
   supports it; some models fix the output count.
2. Ask only about ambiguity that materially changes the result: exact text,
   required invariants, intended use, or missing source image. Do not ask for
   routine details that can be reasonably inferred.
3. Respect a user-selected model. Otherwise apply the maintainable heuristics in
   `references/model-routing.md`. Run `models` when availability matters; a
   network query is not required for every job.
4. Analyze the request, then form one complete final prompt. Keep provider/model
   controls in CLI flags or `--param`, not in prose. If the prompt is already a
   complete spec, pass `--no-augment` to avoid double processing.
5. Prefer local paths for source and reference images. The CLI uploads each local
   PNG, JPEG, or WebP up to 20 MiB through `/image-inputs` before creating the
   job. HTTPS URLs remain supported. Label every input with its role (`source`,
   `reference`, `mask`, `style`, `character`). Masks must be PNG.
6. Run `generate` or `edit`, save outputs deliberately, then read/inspect the
   generated images. The gateway can return more outputs than requested.
7. Iterate one factor at a time: change the prompt, a provider parameter, an
   input role, or the model, then inspect again.

See `references/cli.md` for commands, `references/prompting.md` for prompt
construction, and `references/gateway-api.md` for the stable gateway contract.
Read `references/sample-prompts.md` only when a concrete generation or edit
template would help form the final prompt.

## Basic commands

```bash
uv run scripts/image_gen.py models
uv run scripts/image_gen.py generate --prompt "Editorial photo of a red umbrella" --out output/umbrella
uv run scripts/image_gen.py edit --input source=./photo.jpg --prompt "Replace only the sky with a clear dawn sky" --out output/edit
```

Use `--dry-run` to validate local files and print the upload plan plus job
payload without reading image bytes or requiring a key. Never put API keys in a
prompt, command output, or source file.