Gpt image

gpt-image — image generation via OpenClaw (gpt-image-2)

Wraps openclaw infer image so Claude Code can produce and inspect images. Image generation runs through OpenClaw's OpenAI provider on the Codex/ChatGPT OAuth subscription (transport=codex-responses), so it consumes subscription quota, not OPENAI_API_KEY billing.

All commands go through the helper:

~/.claude/skills/gpt-image/scripts/oc-image.sh <generate|edit|describe|batch|check> [args]

The helper auto-injects sensible defaults (--model openai/gpt-image-2, --json, and an --output path), retries once on transient failures, and prints clean results.

Workflow (follow this every time)

1. Preflight once per session: run oc-image.sh check. If it warns "not authenticated", tell the user to run (themselves, interactively): openclaw models auth login --provider openai and stop until done.
2. Write a strong prompt (see Prompt tips). Prefer English text inside the image — label rendering is much more accurate in English.
3. Generate / edit with the helper. Choose an explicit --output path the user will want (e.g. ./architecture.png); otherwise it defaults to ./oc-image-<timestamp>.png.
4. Always Read the produced PNG to view it, verify the text/layout came out right, then report the saved path to the user.
5. If labels are garbled or the layout is off, iterate: tighten the prompt (fewer words per box, explicit positions) and regenerate, or edit the existing file.

Commands

Generate (text → image):

~/.claude/skills/gpt-image/scripts/oc-image.sh generate \
  --prompt "PROMPT TEXT" \
  --output ./out.png \
  --size 1536x1024 \
  --output-format png        # png | jpeg | webp
# optional: --background transparent   --count 2   --aspect-ratio 16:9

With --count N (N>1) openclaw writes suffixed files out-1.png, out-2.png, … and prints one IMAGE line per file.

Edit (image(s) + prompt → image). Repeat --file for multiple inputs:

~/.claude/skills/gpt-image/scripts/oc-image.sh edit \
  --file ./in.png --prompt "make the background transparent, keep the logo" \
  --output ./edited.png

Describe (image → text). Uses a vision model, not gpt-image-2:

~/.claude/skills/gpt-image/scripts/oc-image.sh describe --file ./in.png \
  --prompt "Describe this in one sentence."

Batch (many prompts → many images). Prompts come from a file (one per line; blank lines and # comments skipped) and/or repeated --prompt. Runs sequentially, continues on error, and writes manifest.tsv in the output dir:

# from a file
~/.claude/skills/gpt-image/scripts/oc-image.sh batch \
  --prompts-file ./prompts.txt --outdir ./renders \
  --size 1536x1024 --output-format png      # passthrough applies to every image

# inline prompts
~/.claude/skills/gpt-image/scripts/oc-image.sh batch \
  --prompt "a red circle icon" --prompt "a green triangle icon" --outdir ./icons

# prompts from stdin
printf 'icon one\nicon two\n' | ~/.claude/skills/gpt-image/scripts/oc-image.sh batch --prompts-file -

Name an output file by prefixing a line with stem<TAB>prompt (otherwise files are 001.png, 002.png, …). Default --outdir is ./oc-batch-<timestamp>/. Passthrough flags (applied to every image): --size --output-format --background --aspect-ratio --resolution --model --timeout-ms. Batch is one image per prompt — --count is ignored here (use generate for multiples).

On success generate/edit/batch print one line per image: IMAGE\t<path>\t<WxH> — Read each <path>. describe prints the description text. After a batch, read <outdir>/manifest.tsv for the stem→status→path→prompt map.

Prompt tips (what makes gpt-image-2 produce good results)

• English labels only inside the image; correct spelling is far more reliable.
• For diagrams / architecture figures: state the layout explicitly — e.g. "three stacked horizontal layers", "circular clockwise loop of 6 stages", "left-to-right pipeline of N boxes connected by arrows". Name each box's exact label and the arrow directions/labels. Add "crisp, correctly-spelled, readable sans-serif labels, flat vector, white background, no clutter".
• Keep per-box text short (2–5 words). Long sentences inside boxes get garbled.
• Specify palette and style ("blue/teal/orange flat enterprise infographic", or "dark navy blueprint with glowing cyan lines") for consistent results.
• The provider auto-expands the prompt (revisedPrompt, echoed to stderr) — useful to see how it interpreted the request when iterating.

Options & sizes

• Sizes (OpenAI provider): 1024x1024, 1536x1024, 1024x1536, 2048x2048, 2048x1152, 3840x2160, 2160x3840. Use landscape 1536x1024 for most diagrams.
• Output formats: png (default), jpeg, webp. Backgrounds: transparent, opaque, auto (transparent only with png/webp).
• Image models available: gpt-image-2 (default), gpt-image-1.5, gpt-image-1, gpt-image-1-mini. Override via --model openai/<id> or env OC_IMAGE_MODEL.

Known limitations

• Exact pixel size is NOT honored under the Codex OAuth backend — --size acts as an aspect/scale hint and the returned image often has different dimensions (e.g. asking 2048×2048 may return ~1254×1254). For exact dimensions, use API-key access instead: export OPENAI_API_KEY=... (then the request routes through the OpenAI Images API). The subscription path is fine for everyday use.
• Very dense diagrams (many tiny labels) may still produce a few misspelled or squeezed labels — split into fewer boxes or iterate.
• Cost: image generation draws on the ChatGPT/Codex subscription quota (check with openclaw models status); a handful of images is a tiny fraction of the 5-hour window. No per-image API charge while on OAuth.

Troubleshooting

• not authenticated → openclaw models auth login --provider openai.
• Unknown model: openai/gpt-image-2 on describe → that's a vision task; the helper already uses openai/gpt-5.5 for describe (override with OC_VISION_MODEL).
• openclaw: command not found → npm install -g openclaw@latest (the helper also tries to add the npm global bin to PATH automatically).