gpt-image-2 - SKILL.md Agent Skill

name: gpt-image-2 description: Use when building or editing a Nebula graph containing a gpt-image-2-* node (generate, edit, fal-generate, fal-edit) or when the user asks to use gpt-image-2 inside Nebula. Covers Nebula-specific node IDs, param names, and UI wiring. For prompting craft, see the global gpt-image-2 skill.

GPT Image 2 — Nebula Integration

OpenAI's gpt-image-2 (released 2026-04-21, snapshot gpt-image-2-2026-04-21) is available in Nebula via four nodes.

Node matrix

Node ID	Path	BYOK key	When to pick
`gpt-image-2-generate`	OpenAI direct	`OPENAI_API_KEY`	Text→image, full streaming previews, need partial frames in canvas
`gpt-image-2-edit`	OpenAI direct	`OPENAI_API_KEY`	Image edit / inpainting with up to 10 reference images + optional mask
`gpt-image-2-fal-generate`	FAL proxy	`FAL_KEY`	Text→image, no OpenAI org-verification needed, pay-per-image
`gpt-image-2-fal-edit`	FAL proxy	`FAL_KEY`	Edit via FAL

All four use executionPattern: "stream". Partial previews render in the canvas as they arrive.

OpenAI-direct params (generate + edit)

Param	Values	Default	Notes
`size`	`auto`, `1024x1024`, `1536x1024`, `1024x1536`, `2048x2048`, `2048x1152`, `3840x2160`, `2160x3840`	`auto`	4K sizes cost significantly more
`quality`	`auto`, `low`, `medium`, `high`	`auto`	~$0.006 / $0.053 / $0.211 at 1024²
`output_format`	`png`, `jpeg`, `webp`	`png`
`output_compression`	0–100	90	Only applied when format != png
`moderation`	`auto`, `low`	`auto`	`low` is less restrictive

Edit node also accepts: images (Image, multiple, up to 10), mask (alpha-channel PNG, same size + format as first image).

FAL-routed params

FAL uses slightly different naming: image_size (not size), num_images (not n). Same sizes and quality values, but num_images maxes at 4 (FAL cap). Otherwise identical semantic surface.

What's NOT supported (and Nebula does not expose)

background: transparent — v1 supported it, v2 does not. If a user asks for transparent backgrounds, suggest chaining a remove-background node downstream instead.
input_fidelity — gpt-image-2 always processes inputs at high fidelity; the param must be omitted.
n > 1 — OpenAI rejects n > 1 when stream=true, and our nodes always stream. The n/Count param is not exposed on the OpenAI-direct nodes. Run the node multiple times for multiple images. (FAL nodes expose num_images because FAL may handle this differently — verify during UAT.)
partial_images (Preview Frames) — removed from the OpenAI-direct node defs. Small/fast generations rarely emit partials (OpenAI docs: "you may not receive the full number of partial images you requested"), and we defaulted to none anyway. FAL nodes still expose it for now pending UAT.

Cost guidance

Draft iteration → quality: low at 1024x1024 ≈ $0.006 each.
Hero asset → quality: high at chosen aspect ratio.
4K → token cost scales roughly with pixel count; use sparingly.
Batch API (50% off) is not wired up in Nebula yet — handle in a later phase.

When editing a prompt for this node

If the user asks Claude to write or refine a prompt that will feed into gpt-image-2-*, consult the global gpt-image-2 skill for prompting craft (5-slot template, anti-slop rules, text-rendering guidance).

Known surprises

OpenAI direct requires Organization Verification — if the user gets "org isn't verified" errors, direct them to https://platform.openai.com/settings/organization/general.
FAL path uses different param names: image_size not size, num_images not n.
Mask is prompt-guided — the prompt must describe the full desired image, not just the masked region.
Up to 10 reference images in the edit endpoint; the mask applies to the first image when supplied.
Streaming: enable partial_images (default 2) to see preview frames in the canvas as the image renders. Set to 0 for final-only.
SSE event namespaces differ by endpoint: /v1/images/generations emits image_generation.completed / image_generation.partial_image, but /v1/images/edits emits image_edit.completed / image_edit.partial_image. Our parser accepts both. OpenAI's docs don't highlight this distinction — discovered during UAT 2026-04-22.