Image Generation

Generate real images from a text description — no paid API key. Two free helper scripts live in this skill's directory, plus a subscription-backed Codex backend for top quality; pick per the rules below. After generating, always open <path> (pops it up in Preview — the user wants to peek at every image) and Read the file so it's shown inline too.

Skill dir: ~/.claude/skills/image-gen/ — comfy_gen.py (local ComfyUI) and pollinations_gen.py (pollinations.ai).

Pick a backend

Check ComfyUI first — it's local, private, unlimited, and free:

curl -s -m 3 -o /dev/null -w "%{http_code}" http://127.0.0.1:8188/system_stats

• 200 → ComfyUI is up. Prefer it for: privacy/offline, batches, square or portrait images (~512–768px native, SD1.5). ~25s per 512² image on this Mac (MPS). Use comfy_gen.py.
• Not 200, OR the user wants a large/widescreen/modern-photoreal image → pollinations. Handles arbitrary sizes (e.g. 1024×576 heroes) and needs no local setup. Requires internet. Use pollinations_gen.py. (Currently serves the sana model; the --model flag is ignored by their API right now — verify the live list with curl -H "User-Agent: Mozilla/5.0" https://image.pollinations.ai/models.)

When unsure, pollinations is the safer default for anything that has to look polished or non-square. ComfyUI wins when local/private/no-rate-limit matters.

• Quality is the whole point (final hero art, logo, text-in-image, photoreal people) → Codex gpt-image-2. Runs on the user's ChatGPT Plus subscription via the local Codex CLI — no API key, but each image turn burns Plus Codex usage 3–5× faster than a normal turn (~10–25 images per 5h window, shared with actual Codex coding use). Finals only; never for drafts, batches, or placeholders — iterate on the free backends first, then do one Codex pass.

Generate

ComfyUI (local, square/portrait):

python3 ~/.claude/skills/image-gen/comfy_gen.py "PROMPT" \
  --out /tmp/<meaningful-name>.png --w 512 --h 512 --steps 25 --seed 42
# add --neg "things to avoid"; --ckpt to pick a checkpoint

Pollinations (any size, widescreen, photoreal):

python3 ~/.claude/skills/image-gen/pollinations_gen.py "PROMPT" \
  --out /tmp/<meaningful-name>.jpg --w 1600 --h 640 --seed 7
# serves the sana model (--model ignored). The helper auto-handles aspect:
# Sana renders 512² square and would STRETCH non-square / LETTERBOX large
# requests, so the helper renders a modest square then crops+upscales to your
# w×h — undistorted, full-bleed, any aspect. Pass --raw to bypass (may stretch).

Codex gpt-image-2 (subscription, best quality — finals only):

codex exec --skip-git-repo-check 'Use your image generation tool: PROMPT — style/detail notes. Save the image to /tmp/<meaningful-name>.png and output only that path. Do nothing else.' < /dev/null

ALWAYS append < /dev/null. Without it, codex exec blocks forever on "Reading additional input from stdin…" in non-interactive/agent shells (stdin never sends EOF) — this is the real cause of the so-called "hang," not OpenAI. With stdin closed it returns in seconds–minutes. The saved path is the last non-log line of stdout. Sizes: 1024x1024, 1024x1536, 1536x1024 (model may render larger and downscale via sips — exact pixels not guaranteed). Raw outputs also land in ~/.codex/generated_images/<session>/. One image ≈ 45K tokens of Plus usage; do NOT loop this in a batch.

Text-in-image strength (Codex only): gpt-image-2 renders lots of text accurately — it's the right backend for real infographics, charts, diagrams, slides, posters, and anything with labels/numbers. (The free backends — Sana/pollinations and SD1.5/ComfyUI — garble text; don't use them when words must be legible.) To get a true data infographic rather than generic "infographic-looking" art, spell out every text string verbatim in quotes AND say where each goes — header, a stat row of "NUMBER / LABEL" cards, named sections, per-bar chart labels ("Adam Detrick 399"), timeline rows, footer. Verified Jun 16 2026: a single image cleanly rendered ~30+ text elements across a multi-section portrait layout (header, 4 stat cards, a 6-item feature list, a 5-bar labeled chart, two big callouts, a 4-item list, a footer strip) with correct spelling. It handles 2:3 portrait well (rendered 864×1821). Still spot-check spelling at full size — it occasionally drops a character on dense small text; code-syntax strings with parens/symbols are the most fragile, so phrase labels as plain readable words where you can.

Texture gotcha: gpt-image-2 piles on crusty over-sharpened micro-texture ("overcooked" — the user's words) unless prompted away. Append something like "smooth painterly rendering, clean forms, soft atmospheric depth, no oversharpened texture, no HDR grit" to every gpt-image-2 prompt that isn't deliberately gritty.

Concurrency gotcha: never run two codex exec calls at once — concurrent execs wedge each other indefinitely (seen Jun 11 2026: parallel batch hung on every turn while a serial one proceeded). Strictly one at a time.

Hang gotcha — ROOT CAUSE FOUND (Jun 16 2026): the "wedges forever" failure is almost always codex exec blocking on an open stdin ("Reading additional input from stdin…"), NOT OpenAI's stream dying. In this agent harness stdin is an open pipe that never EOFs, so codex waits indefinitely. Fix: always run with < /dev/null (or stdin=subprocess.DEVNULL from Python). Verified Jun 16: a trivial text exec that "hung" past 150s with open stdin returned in 9s with < /dev/null; the infographic image then generated cleanly in well under a minute. Still keep a watchdog (subprocess.run(cmd, timeout=600, stdin=subprocess.DEVNULL)) as a backstop for genuine stream death, but redirecting stdin is the actual cure. Don't >/dev/null the output — you need the path/errors.

Aspect-ratio gotcha (pollinations only): never trust a raw non-square pollinations request — Sana stretches it. The helper fixes this for you; just pass the --w/--h you actually want. For HTML heroes you can also request a square and let CSS object-fit:cover shape it. ComfyUI honors exact dimensions natively (no stretch) but SD1.5 degrades past ~768px and on very wide frames.

The helper scripts print the saved path on success. A fixed --seed makes results reproducible; omit/vary it for variety (Codex has no seed control). Name output files meaningfully (subject, not image1), and after generating, open <path> (Preview popup — default, every image) then Read the file to display it inline.

Writing the prompt

Be specific and visual: subject + setting + style + lighting + detail level. e.g. "a red fox curled asleep in autumn leaves, golden hour, shallow depth of field, photorealistic, highly detailed". Name an art style when it matters (watercolor, flat vector, isometric 3D, oil painting, line art). For ComfyUI/SD1.5, the default negative prompt already filters common artifacts; add specifics (e.g. --neg "extra fingers, text, signature") when needed.

If the user only wants creative prompt ideas as text (not a rendered image), that's the image-prompts skill, not this one.

Failure handling

• ComfyUI submit fails / errors on server → fall back to pollinations.
• Pollinations 403 → the helper already sends a browser User-Agent; a persistent 403/empty body means rate-limited or down, so fall back to ComfyUI (if up) or tell the user.
• Pollinations 402 Payment Required (first seen Jun 11, 2026 — generation 402s even though /models still answers) → their free anonymous access may be ending or temporarily gated. Fall back to ComfyUI; if this persists across days, treat pollinations as dead and consider replacing this backend.
• Don't silently retry more than the scripts already do; report which backend you used.