菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
FAL.ai model catalog access — image, video, audio, 3D, LoRA training, and more via fal-ai/ and bytedance/ and wan/ endpoints. Activate when the user configures any nebula node that routes through FAL (kling, sora, veo, flux, seedance, wan, luma, pixverse, ltx, moonvalley, hunyuan3d, recraft, nano-banana via FAL, and 30+ others), when the user mentions a fal-ai/ slug, or when they ask which FAL model to use for a task. Entry point for 160 curated per-model skills + category overviews at `.agents/skills/fal/`.
FAL (fal.ai) is Nebula's all-purpose generation gateway — one API key fronts 52 ready-to-use nodes for image, video, 3D, and audio generation (FLUX, GPT Image 1.5/2, Seedream, Ideogram 4 (gen + edit/remix/reframe/replace-background/character/upscale), Recraft, Sora 2, Kling, Wan, Luma Ray 2, PixVerse, LTX, Seedance, Hunyuan3D, Meshy, Stable Audio 2.5, ACE-Step, MMAudio V2, Demucs, Clarity Upscaler, SeedVR2 Video Upscale) plus the catch-all `fal-universal` slug node. Activate when the user configures any FAL node — flux-1-1-ultra, flux-2-pro, flux-schnell, fast-sdxl, flux-kontext, flux-fill-inpaint, gpt-image-1-5, gpt-image-1-5-edit, gpt-image-2-fal-generate, gpt-image-2-fal-edit, seedream-4-5, ideogram-v4, ideogram-edit, ideogram-remix, ideogram-reframe, ideogram-replace-background, ideogram-character, ideogram-upscale, recraft-v4-raster, recraft-v4-svg, sora-2, kling-v2-1, kling-v3, kling-o3, wan-2-6-t2v, wan-2-6-i2v, wan-2-6-r2v, luma-ray2-t2v, luma-ray2-i2v, luma-ray2-flash-modify, pixverse-v4-5, ltx-video
Anthropic (Claude) Messages API in Nebula — a single text-gen node (Claude / claude-chat) that turns a text prompt plus optional image(s) into streamed text, with system prompt, temperature/top_p/stop sampling, and extended-thinking. Use for captioning, summarizing, translating, rewriting, reasoning, and drafting prompts to feed other media nodes. Activate when the user configures the `claude-chat` node, wires a Claude/Anthropic node into a graph, or asks about Anthropic/Claude in Nebula. Sourced from the official Anthropic Messages API (platform.claude.com/docs/en/api/messages) and the Nebula audit guide docs/api-guides/anthropic.md on 2026-06-04.
ElevenLabs audio API in Nebula — text-to-speech, sound-effects-from-text, speech-to-speech voice changing, audio isolation (denoise), translated dubbing, and speech-to-text transcription, all as audio-gen nodes. Activate when the user configures any elevenlabs-tts, elevenlabs-sfx, elevenlabs-sts, elevenlabs-isolation, elevenlabs-dubbing, or elevenlabs-stt node, or asks about ElevenLabs in Nebula. Sourced from the official ElevenLabs API reference (elevenlabs.io/docs/api-reference) cross-checked against backend/handlers/elevenlabs.py + backend/data/node_definitions.json, and the Nebula audit guide docs/api-guides/elevenlabs.md, on 2026-06-05.
Best practices for prompting Google's Gemini-family APIs — Gemini text models, Imagen, Nano Banana (Gemini 3.1 Flash Image), and Veo video. Activate when writing prompts, building graphs, or configuring nodes that target any Gemini model, when editing files importing `google-genai` / `@google/genai`, or when the user asks about Imagen, Nano Banana, or Veo. Routes to topic files for deep guidance. Sourced from ai.google.dev with fetch dates noted — verify against the live docs for anything that smells stale.
Higgsfield video generation in Nebula — text-to-video and image-to-video short cinematic clips via the DoP ("Depth of Presence") model plus Kling v2.1 Pro and Seedance v1 Pro guest models, with duration and aspect-ratio control. Activate when the user configures the `higgsfield` node or asks about Higgsfield in Nebula. The node is video-output only (no Soul images, Speak lipsync, or motion-camera presets are wired). Sourced 2026-06-04 from the official Higgsfield API docs (docs.higgsfield.ai) plus the Nebula audit guide docs/api-guides/higgsfield.md.
| 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. |
OpenAI's gpt-image-2 (released 2026-04-21, snapshot gpt-image-2-2026-04-21) is available in Nebula via four nodes.
| 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.
| 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 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.
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.quality: low at 1024x1024 ≈ $0.006 each.quality: high at chosen aspect ratio.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).
image_size not size, num_images not n.partial_images (default 2) to see preview frames in the canvas as the image renders. Set to 0 for final-only./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.