gpt-image-edit（GPT 图像编辑）

GPT Image Edit — Pro Pack on RunComfy

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

OpenAI GPT Image 2 — /edit endpoint (ChatGPT Images 2.0 image-to-image) on the RunComfy Model API. Strongest in its class at preserving identity through targeted edits and rewriting embedded text in any script (Latin, kana, CJK, Cyrillic, Arabic).

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

When to pick this model (vs siblings)

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

openai/gpt-image-2/edit

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

How to invoke

Single-ref preservation edit:

runcomfy run openai/gpt-image-2/edit \
  --input '{
    "prompt": "Keep the person'\''s face, pose, and brand mark unchanged. Replace the background with a soft warm-grey studio sweep and a gentle floor shadow.",
    "images": ["https://.../portrait.jpg"]
  }' \
  --output-dir <absolute/path>

Multilingual text rewrite (preserve everything except the headline):

runcomfy run openai/gpt-image-2/edit \
  --input '{
    "prompt": "Keep the photograph, layout, and brand mark exactly as in the input. Replace only the in-image headline. The new headline reads \"今日のおすすめ\" in bold Japanese kana, same position and font weight as before.",
    "images": ["https://.../poster-en.jpg"]
  }' \
  --output-dir <absolute/path>

Multi-ref composition:

runcomfy run openai/gpt-image-2/edit \
  --input '{
    "prompt": "Compose subject from image 1 into the room from image 2. Match the lighting and color palette of image 2. Keep image 1 subject identity (face, pose, clothing) unchanged.",
    "images": ["https://.../subject.jpg", "https://.../room.jpg"]
  }' \
  --output-dir <absolute/path>

Prompting — what actually works

Lead with preservation goals. Always: "Keep [face / pose / clothing / brand / framing] unchanged." Then state the change. The model honors what's stated up front.

Multilingual text — quote the characters, name the script. "the headline reads \"コーヒー\" in bold Japanese kana", "the label says \"АРОМА\" in Cyrillic, white on black", "the right-margin caption reads \"تخفيض\" in Arabic right-to-left". Don't paraphrase — quote.

Directional language for spatial edits. Concrete spatial scopes work: "move the headline from top-right to bottom-center", "remove the leftmost object only", "replace the watermark in the bottom-right corner".

Multi-ref numbering. When passing multiple images, refer to them by number: "subject from image 1, lighting from image 2, color palette from image 3". The model routes cues correctly.

Use size: "auto" to preserve input ratio. Only override when the edit explicitly changes framing (e.g. cropping a 16:9 to 1:1).

Anti-patterns:

• Long compound edit instructions ("change A and B and C and D") → drift increases per added scope.
• Missing preservation goals → model subtly rewrites the face / brand / framing.
• Paraphrasing in-image text instead of quoting it → text comes out different.
• Asking for size outside the 3 fixed values + auto → 422.

Where it shines

Sample prompts (verified to produce strong results)

Background swap with full preservation (page example):

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

Multilingual variant:

Keep the photograph, layout, lighting, and brand mark exactly as in the
input. Replace only the in-image headline.
The new headline reads "コーヒー" in bold Japanese kana, same position
and font weight as before.

Multi-ref composition:

Compose subject from image 1 into the kitchen from image 2.
Match the warm window light and color palette of image 2.
Keep subject identity (face, pose, clothing) from image 1 unchanged.

Limitations

• size: 3 fixed values + auto — anything else 422s.
• images: up to 10 — first is primary, rest are auxiliary cues.
• Long compound prompts drift — split into multiple passes when needed.
• For batch consistency across many SKU images, Nano Banana Edit (up to 20) is better.
• Photorealism on portraits — Nano Banana Pro wins head-to-head.

Exit codes

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

How it works

The skill invokes runcomfy run openai/gpt-image-2/edit with a JSON body matching the schema. The CLI POSTs to https://model-api.runcomfy.net/v1/models/openai/gpt-image-2/edit, polls the request, fetches the result, and downloads any .runcomfy.net/.runcomfy.com URL into --output-dir. Ctrl-C cancels the remote request before exit.

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.