gpt-image-2（GPT Image 2 图像生成）

GPT Image 2 — Pro Pack on RunComfy

runcomfy.com · Text-to-image · Edit · GitHub

OpenAI GPT Image 2 (ChatGPT Images 2.0) hosted on the RunComfy Model API — no OpenAI key, async REST.

npx skills add agentspace-so/runcomfy-skills --skill gpt-image-2 -g

When to pick this model (vs siblings)

GPT Image 2's distinct strength is directive precision: it follows multi-element prompts, layout cues, and embedded-text instructions more reliably than its peers. Pick it when what's on the canvas matters more than how stylized it looks.

If the user explicitly asked for GPT Image 2 / ChatGPT Image 2 / Image 2, route here regardless — don't second-guess the model choice.

Prerequisites

1. RunComfy CLI — npm i -g @runcomfy/cli
2. RunComfy account — runcomfy login opens a browser device-code flow.
3. CI / containers — set RUNCOMFY_TOKEN=<token> instead of runcomfy login.

Endpoints + input schema

Two endpoints, same model.

openai/gpt-image-2/text-to-image

openai/gpt-image-2/edit

size=auto on edit preserves the input aspect ratio — strongly recommended unless the edit explicitly changes framing.

How to invoke

Text-to-image:

runcomfy run openai/gpt-image-2/text-to-image \
  --input '{"prompt": "<user prompt>", "size": "1024_1536"}' \
  --output-dir <absolute/path>

Edit (single ref):

runcomfy run openai/gpt-image-2/edit \
  --input '{
    "prompt": "<edit instruction>",
    "images": ["https://..."]
  }' \
  --output-dir <absolute/path>

Edit (multi-ref, up to 10):

runcomfy run openai/gpt-image-2/edit \
  --input '{
    "prompt": "compose subject from image 1 into the room from image 2; match the lighting of image 2",
    "images": ["https://...subject.jpg", "https://...room.jpg"]
  }' \
  --output-dir <absolute/path>

The CLI submits, polls every 2s until terminal, then downloads any *.runcomfy.net / *.runcomfy.com URL from the result into --output-dir. Stdout is the result JSON. Stderr is progress.

For pipe-friendly usage:

runcomfy --output json run openai/gpt-image-2/text-to-image \
  --input '{"prompt":"..."}' --no-wait | jq -r .request_id

Prompting — what actually works

These are model-specific patterns that empirically improve output quality. Apply to text-to-image and edit alike.

Be explicit on subject + setting + mood. "A close-up of a matte ceramic water bottle on warm linen, soft window light, neutral background" — three concrete directives — beats "nice product photo of a bottle".

Quote embedded text exactly. Keep it short. GPT Image 2 is the strongest text-rendering model in this class, but only when you put the literal characters in quotes. Long blocks of text degrade. For multilingual text, name the script: "Japanese kana", "Cyrillic", "Arabic right-to-left".

Use compositional cues directly. "rule of thirds", "close-up", "aerial view", "centered subject", "shallow depth of field" — these have learned-meaning to the model.

Iterate one attribute at a time. When refining, change one thing per iteration (lighting OR background OR pose OR text) and keep the rest of the prompt verbatim. The model holds composition stable across iterations when only one knob moves.

Don't conflict instructions. "no text" + "the word 'AQUA+' on the label" is incoherent — the model will pick one and you don't control which.

Don't pile up styles. "ukiyo-e + watercolor + 8K + cinematic + minimalist" cancels out. Pick one or two style anchors max.

For the edit endpoint specifically:

• State preservation goals. "keep the person's pose and face identity unchanged", "keep the brand mark and typography on the package", "keep the overall framing". The model needs to know what NOT to change.
• Use directional language for spatial edits. "Move the headline from top-right to bottom-center", not "reposition the headline".
• Multi-ref: number the images in the prompt — "subject from image 1, lighting and background from image 2" — and the model will route the cues correctly.

Where it shines

Sample prompts (verified to produce strong results)

Text-to-image — product hero:

A minimal hero product still life: a matte ceramic water bottle on warm linen,
soft window light, the word "AQUA+" in clean sans-serif on the label,
subtle rim highlights, e-commerce ready, 8K detail, neutral background

Text-to-image — multilingual signage:

A small Tokyo café storefront at dusk, warm interior glow,
the sign reads "コーヒー" in bold Japanese kana on a wooden plaque,
shallow depth of field, rule of thirds, cinematic

Edit — background swap with preservation:

Turn the background into a bright minimal white-to-soft-gray studio sweep
with gentle floor shadow; add a large headline in-image that reads
"OPEN STUDIO" in a bold clean sans-serif, high contrast, centered;
keep the main person or product, pose, and face identity unchanged

Limitations

• Only 3 fixed sizes on text-to-image (and the same 3 + auto on edit). Extreme aspect ratios are auto-resized to the nearest supported one.
• Prompt length ~ a few thousand tokens. Long blocks of embedded text degrade output.
• Edit's multi-image support is "guidance from up to 10 refs", not ControlNet-style stacks. The first image is treated as the primary; the rest provide auxiliary cues.
• Photorealism on portraits is not its strongest suit — Nano Banana Pro wins that head-to-head.

Exit codes

The runcomfy CLI uses sysexits-style codes:

Full reference: docs.runcomfy.com/cli/troubleshooting.

How it works

1. The skill invokes runcomfy run openai/gpt-image-2/<endpoint> with a JSON body matching the schema above.
2. The CLI POSTs to https://model-api.runcomfy.net/v1/models/openai/gpt-image-2/<endpoint> with the user's bearer token.
3. The Model API returns a request_id; the CLI polls GET.../requests/<id>/status every 2 seconds.
4. On terminal status, the CLI fetches GET.../requests/<id>/result and downloads any URL whose host ends with .runcomfy.net or .runcomfy.com into --output-dir. Other URLs are listed but not fetched.
5. Ctrl-C while polling sends POST.../requests/<id>/cancel so you don't get billed for GPU you stopped.

What this skill is not

Not a direct OpenAI API client. Not a capability grant — depends on a working RunComfy account. Not multi-tenant.

Security & Privacy

• Token storage: runcomfy login writes the API token to ~/.config/runcomfy/token.json with mode 0600 (owner-only read/write). Set RUNCOMFY_TOKEN env var to bypass the file entirely in CI / containers.
• Input boundary: the user prompt is passed as a JSON string to the CLI via --input. The CLI does NOT shell-expand the prompt; it transmits the JSON body directly to the Model API over HTTPS. No shell injection surface from prompt content.
• Third-party content: image / mask / video URLs you pass are fetched by the RunComfy model server, not by the CLI on your machine. Treat external URLs as untrusted; image-based prompt injection is a known risk for any image-edit / video-edit model.
• Outbound endpoints: only model-api.runcomfy.net (request submission) and *.runcomfy.net / *.runcomfy.com (download whitelist for generated outputs). No telemetry, no callbacks.
• Generated-file size cap: the CLI aborts any single download > 2 GiB to prevent disk-fill from a malicious or runaway model output.