Image Generation

The image-generate skill turns a text prompt into an image file on disk, fully from the command line, so it drops into any pipeline. It has swappable provider backends (selected with --provider) so the capability isn't locked to one vendor's pricing or availability.

The deterministic engine is a stdlib-only script, plugins/nolte-media/skills/image-generate/scripts/image_generate.py; the skill is the operator-facing wrapper.

Providers at a glance

--provider	Credentials needed	Free?	Output licence	Use it for
cloudflare (default)	CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID	Yes, real free tier (no card)	FLUX.1-schnell, Apache-2.0; you own the output	Blog and production images
pollinations	none (optional POLLINATIONS_API_TOKEN)	Yes, auth-free	No explicit licence (defers to the model); public-feed default	Quick throwaway images
gemini	GEMINI_API_KEY	No, requires billing	Commercial use allowed	Only if you already pay for it

Recommendation: use cloudflare for anything you publish. It's free, carries no watermark or public feed, and FLUX.1-schnell is Apache-2.0, so the generated images are clearly yours to use. pollinations is fine for disposable images but grants no explicit output licence; gemini needs a billed project (its free-tier quota for this image model is zero).

Which token does each provider need?

Cloudflare (default, recommended)

Cloudflare Workers AI runs FLUX.1-schnell on a real recurring free tier (10,000 neurons/day, no credit card). You need a token and your account id:

1. Create a free account at https://dash.cloudflare.com/sign-up (no card required).
2. In the dashboard go to AI → Workers AI → "Use REST API" (REST is Representational State Transfer).
3. Click "Create a Workers AI API Token". Cloudflare pre-fills the right permissions:
4. Account · Workers AI · Read
5. Account · Workers AI · Edit (this one is required, because generating an image is a run/Edit action)
6. Create the token and copy it (shown only once).
7. Copy your Account ID from the same page.

Then export both (the tool reads them only from the environment, never as a flag):

bash export CLOUDFLARE_API_TOKEN="your_token" export CLOUDFLARE_ACCOUNT_ID="your_account_id"

The free tier resets daily; FLUX.1-schnell costs ~4.8 neurons per 512x512 tile, so 10,000 neurons/day is roughly hundreds of images.

Pollinations (auth-free)

No token is required. Optionally set POLLINATIONS_API_TOKEN to remove the watermark that anonymous requests carry. The first run prints a one-time disclaimer (public feed, no explicit output licence) that you must acknowledge.

Gemini (billing required)

Set GEMINI_API_KEY from https://aistudio.google.com/apikey. Note that gemini-2.5-flash-image isn't on the free tier (its free-tier quota is 0); the project must have billing enabled or every call returns a billing-required error.

Generate an image

The default provider is cloudflare, so no --provider is needed for it:

```bash

via the Taskfile (pass script args after --)

task image:generate -- --prompt "a minimalist teal fox icon, flat" --out fox.jpg

or directly

python3 "${CLAUDE_PLUGIN_ROOT}/skills/image-generate/scripts/image_generate.py" \ --prompt "a minimalist teal fox icon, flat" --out fox.jpg ```

Switch backends with --provider:

bash python3 "${CLAUDE_PLUGIN_ROOT}/skills/image-generate/scripts/image_generate.py" \ --provider pollinations --prompt "a tree, comic style" --out tree.jpg --accept-data-policy

Every run writes the image plus a <image>.meta.json sidecar recording provider, model, source, prompt, timestamp, and mime_type.

Useful flags: --from-prompt-doc <doc> --variant light|dark (render a graphic-prompt-generator document), -n <N> (several images of the same prompt), --seed, --width/--height, --force (overwrite an existing file).

Tip

Cloudflare and Pollinations both return JPEG. Use a .jpg target to avoid the extension/MIME-mismatch warning (MIME is Multipurpose Internet Mail Extensions) (the image is still written either way).

Using it in another repo

When the nolte-media plugin is installed in another repository, this capability is reachable there as /nolte-media:image-generate; nothing is copied into the consumer repo. The skill invokes the bundled script through ${CLAUDE_PLUGIN_ROOT}, which resolves to the installed plugin's directory in any context (marketplace install and claude --plugin-dir ./plugins/nolte-media dogfooding), so the same command works everywhere:

bash python3 "${CLAUDE_PLUGIN_ROOT}/skills/image-generate/scripts/image_generate.py" \ --prompt "a teal fox icon, flat" --out fox.jpg

Only the script path is plugin-rooted; data paths (--out, --from-prompt-doc) stay relative to the consumer's working directory. The task image:generate shortcut is claude-shared-local (it runs the repo-relative script outside Claude); consumer repos use the slash command instead.

Fits a pipeline

The tool closes a three-step asset pipeline:

1. graphic-prompt-generator (agent) authors a brand-conformant prompt document under design/prompts/.
2. image-generate renders it: --from-prompt-doc design/prompts/hero.md --variant dark --out hero.jpg.
3. png-to-transparent-svg (agent) turns the result into a vector for an icon or logo.

Exit codes

0 success · 1 runtime error (network, filesystem, malformed response) · 2 usage error · 3 rate limit / quota / billing-required · 4 authentication failure. Nothing is retried automatically on 429.