Image Edit — Pro Pack on RunComfy

runcomfy.com · Nano Banana Edit · GPT Image 2 Edit · Flux Kontext · Z-Image Inpaint · GitHub

Image edit, intent-routed. This skill doesn't lock you to one model — it picks the right edit model in the RunComfy catalog based on what the user actually wants: batch identity-preservation, multilingual text rewrite, single-shot precise edit, or mask-driven region replacement.

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

Pick the right model for the user's intent

User intent	Model	Why
Batch edit 1–20 images consistently (SKU gallery, A/B variants)	Nano Banana Edit	Up to 20 input images per call; locked aspect/resolution for series
Swap background, preserve subject identity	Nano Banana Edit	Strong identity preservation under "keep X unchanged" prompts
Localized object removal / addition with spatial language ("the left object", "upper-right corner")	Nano Banana Edit	Honors directional spatial scope
Multilingual / non-Latin in-image text rewrite (Japanese kana, Cyrillic, Arabic)	GPT Image 2 Edit	Strongest in class for multilingual typography
Multi-reference composition (subject from img1, scene from img2, palette from img3)	GPT Image 2 Edit	Numbered refs route cues correctly
Layout-precise repositioning ("move headline from top-right to bottom-center")	GPT Image 2 Edit	Directional language honored at layout level
Identity preservation across translated headline variants	GPT Image 2 Edit	Same source asset → many language variants, identity stable
Single-shot precise local edit ("she's now holding an orange umbrella")	Flux Kontext Pro	Single-ref single-instruction, high-fidelity preservation
Mask-driven object removal (cables, watermarks, distractions)	Z-Image Turbo Inpaint	Mask-required, strength-tunable, edge-consistent
Mask-driven region replacement (full background swap with mask)	Z-Image Turbo Inpaint	High strength + clean mask = clean replacement
Default if unspecified	Nano Banana Edit	Most flexible, supports both single and batch

The agent reads this table, classifies the user's intent, and picks the matching subsection below.

Prerequisites

1. RunComfy CLI — npm i -g @runcomfy/cli
2. RunComfy account — runcomfy login.
3. CI / containers — set RUNCOMFY_TOKEN=<token>.

---

Route 1: Nano Banana Edit — default for general edit + batch

Model: google/nano-banana-2/edit

Schema

Field	Type	Required	Default	Notes
prompt	string	yes	—	Lead with preservation goals, end with the change.
image_urls	array	yes	—	1–20 publicly-fetchable HTTPS URLs.
number_of_images	int	no	1	1–4 outputs per call.
aspect_ratio	enum	no	auto	auto follows input; lock for batch consistency.
resolution	enum	no	1K	0.5K / 1K / 2K / 4K.
output_format	enum	no	png	png / jpeg / webp.
seed	int	no	—	Reproducibility.
enable_web_search	bool	no	false	Web-grounded edits (extra latency).

Invoke

runcomfy run google/nano-banana-2/edit \
  --input '{
    "prompt": "Keep the subject identity, pose, and clothing unchanged. Convert the background into a rainy neon cyberpunk street.",
    "image_urls": ["https://.../portrait.jpg"]
  }' \
  --output-dir <absolute/path>

Batch (lock aspect + resolution):

runcomfy run google/nano-banana-2/edit \
  --input '{
    "prompt": "Replace the watermark in the bottom-right with the text \"AURA\" in clean white sans-serif. Keep everything else exactly as in the input.",
    "image_urls": ["https://.../sku-1.jpg", "https://.../sku-2.jpg", "https://.../sku-3.jpg"],
    "aspect_ratio": "1:1",
    "resolution": "1K"
  }' \
  --output-dir <absolute/path>

Prompting tips

• Preservation first: "Keep [identity / pose / brand / framing] unchanged." Then state the change.
• Spatial scope: "background only", "the left object", "upper-right quadrant" — concrete locations honored.
• Batch consistency: lock aspect_ratio and resolution across the batch.
• Iterate small: split compound edits into multiple shorter passes.

---

Route 2: GPT Image 2 Edit — multilingual text + multi-ref composition

Model: openai/gpt-image-2/edit

Schema

Field	Type	Required	Default	Notes
prompt	string	yes	—	Edit instruction; lead with preservation.
images	string[]	yes	—	Up to 10 HTTPS URLs. First is primary; rest are auxiliary.
size	enum	no	auto	auto, 1024_1024, 1024_1536, 1536_1024. Only these.

Invoke

Multilingual text rewrite:

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.",
    "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 unchanged.",
    "images": ["https://.../subject.jpg", "https://.../room.jpg"]
  }' \
  --output-dir <absolute/path>

Prompting tips

• Quote in-image text exactly. Name the script for non-Latin: "Japanese kana", "Cyrillic", "Arabic right-to-left".
• Number multi-refs: "subject from image 1, lighting from image 2".
• Directional layout language: "move the headline from top-right to bottom-center", "replace the watermark in the bottom-right".
• size: "auto" preserves input ratio — recommended unless the edit changes framing.

---

Route 3: Flux Kontext Pro — single-shot precise local edit

Model: blackforestlabs/flux-1-kontext/pro/edit

Schema (minimal)

Field	Type	Required	Notes
prompt	string	yes	One declarative edit instruction.
image	string	yes	Single source image URL.
aspect_ratio	enum	no	Pick from supported W:H values.
seed	int	no	Reproducibility.

Single image only — no array. For multi-image flows, use Route 1 (Nano Banana Edit).

Invoke

runcomfy run blackforestlabs/flux-1-kontext/pro/edit \
  --input '{
    "prompt": "Keep the person'\''s face, pose, and clothing unchanged. Add an orange umbrella in her left hand and a slight smile.",
    "image": "https://.../portrait.jpg"
  }' \
  --output-dir <absolute/path>

Prompting tips

• One declarative instruction. "She is now holding an orange umbrella and smiling" — imperative, single change.
• Preservation first. Lead with "Keep [unchanged elements]" then state the change.
• Iterate small. Compound edits drift on a single pass; split into sequential passes.

---

Route 4: Z-Image Turbo Inpaint — mask-driven precise region edit

Model: tongyi-mai/z-image/turbo/inpainting

Schema

Field	Type	Required	Notes
prompt	string	yes	What to fill / replace; preservation constraints for the unmasked surround.
image	string	yes	Source image URL.
mask_image	string	yes	Grayscale mask URL (white = inpaint, black = preserve).
strength	float	no	0.3–0.6 retouching, 0.7–1.0 full replacement.
control_scale	float	no	0.6–0.9 typical.
aspect_ratio	enum	no	W:H output ratio.
seed	int	no	Reproducibility.

Invoke

Object removal (low strength):

runcomfy run tongyi-mai/z-image/turbo/inpainting \
  --input '{
    "prompt": "Remove overhead cables; preserve rooflines and sky gradient; thin clean sky.",
    "image": "https://.../street.jpg",
    "mask_image": "https://.../cables-mask.png",
    "strength": 0.5,
    "control_scale": 0.8
  }' \
  --output-dir <absolute/path>

Region replacement (high strength):

runcomfy run tongyi-mai/z-image/turbo/inpainting \
  --input '{
    "prompt": "Replace busy backdrop with smooth light gray studio paper; mask background only.",
    "image": "https://.../product.jpg",
    "mask_image": "https://.../bg-mask.png",
    "strength": 0.9
  }' \
  --output-dir <absolute/path>

Prompting tips

• A mask URL is required — grayscale, white = inpaint region, black = preserve. Slight blur on mask edges (1–3px) blends better than sharp binary.
• Strength by intent: 0.3–0.5 for retouching / cleanup, 0.6–0.7 for object replacement with style match, 0.8–1.0 for full-region replacement.
• Name what stays outside the mask in the prompt: "preserve rooflines and sky gradient", "match brick pattern and mortar tone".
• Spatial labels still help even though the mask defines the region: "the left shelf", "upper-right quadrant".

---

Limitations

• Each route inherits its model's limits. Nano Banana: 1–20 inputs, 1–4 outputs. GPT Image 2 Edit: up to 10 refs, 4 fixed sizes. Flux Kontext: single ref. Z-Image Inpaint: mask required.
• No multi-route blending. This skill picks one model per call.
• Brand-specific overrides — if the user named a specific model, route to the corresponding brand skill (gpt-image-edit, flux-kontext, nano-banana-edit) for fuller treatment.

Exit codes

code	meaning
0	success
64	bad CLI args
65	bad input JSON / schema mismatch
69	upstream 5xx
75	retryable: timeout / 429
77	not signed in or token rejected

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

How it works

The skill picks one of Nano Banana Edit / GPT Image 2 Edit / Flux Kontext Pro / Z-Image Turbo Inpaint based on user intent and invokes runcomfy run <model_id> with the matching JSON body. The CLI POSTs to the Model API, 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.