OpenAI (BYOK, direct) in Nebula — text-to-image (GPT Image 1/1.5/mini, GPT Image 2 up to 4K), image edit + inpaint (single image or up to 10 refs, optional mask), text-to-speech (13 voices, 6 formats, voice instructions), Whisper/GPT-4o transcription (text/json/srt/vtt), audio translation to English, and OpenAI Chat (GPT-5.5/5.4 + legacy GPT-4o/4.1) with vision and JSON-object output. Activate when the user configures any gpt-image-1-generate, gpt-image-1-edit, gpt-image-2-generate, gpt-image-2-edit, openai-tts, openai-stt, openai-translate, or gpt-4o-chat node, or asks about OpenAI in Nebula. For gpt-image-2 prompting craft and the FAL-routed gpt-image-2 nodes, see .agents/skills/gpt-image-2/SKILL.md. Sourced from the official OpenAI API docs (platform.openai.com / developers.openai.com) and the Nebula audit guide docs/api-guides/openai.md, verified against the backend handlers on 2026-06-04 (updated 2026-06-10).
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
OpenAI (BYOK, direct) in Nebula — text-to-image (GPT Image 1/1.5/mini, GPT Image 2 up to 4K), image edit + inpaint (single image or up to 10 refs, optional mask), text-to-speech (13 voices, 6 formats, voice instructions), Whisper/GPT-4o transcription (text/json/srt/vtt), audio translation to English, and OpenAI Chat (GPT-5.5/5.4 + legacy GPT-4o/4.1) with vision and JSON-object output. Activate when the user configures any gpt-image-1-generate, gpt-image-1-edit, gpt-image-2-generate, gpt-image-2-edit, openai-tts, openai-stt, openai-translate, or gpt-4o-chat node, or asks about OpenAI in Nebula. For gpt-image-2 prompting craft and the FAL-routed gpt-image-2 nodes, see .agents/skills/gpt-image-2/SKILL.md. Sourced from the official OpenAI API docs (platform.openai.com / developers.openai.com) and the Nebula audit guide docs/api-guides/openai.md, verified against the backend handlers on 2026-06-04 (updated 2026-06-10).
OpenAI Skill
These are the eight OpenAI-direct nodes (all BYOK, one key). For prompting craft on the GPT Image 2 nodes and for the FAL-routedgpt-image-2-fal-* nodes (different provider, FAL_KEY, image_size/num_images naming), defer to .agents/skills/gpt-image-2/SKILL.md — this skill does not duplicate that detail.
When to use
User drags or configures any of: gpt-image-1-generate, gpt-image-1-edit, gpt-image-2-generate, gpt-image-2-edit, openai-tts, openai-stt, openai-translate, gpt-4o-chat.
User wants to turn text → image, edit/inpaint an image, read text aloud, transcribe or translate speech, or run OpenAI chat (GPT-5.x or legacy 4o/4.1, including describing piped-in images).
User asks which OpenAI model/voice/format to pick, what a param does, or why an OpenAI run failed (org verification, English-only translate, missing key).
User asks to chain OpenAI nodes (e.g. image → caption → speech, or audio → subtitles).
Universal rules
Auth + env var. Every node sends Authorization: Bearer <OPENAI_API_KEY>. One key (OPENAI_API_KEY) covers all eight. Set it in .env at repo root or backend/, then restart the backend (.env is read at boot). Missing key → handler raises OPENAI_API_KEY is required before any network call.
Base URL.https://api.openai.com/v1. Endpoints used: /images/generations, /images/edits, /audio/speech, /audio/transcriptions, /audio/translations, /chat/completions. No base-URL override in the handlers.
Execution pattern (confirmed per handler):
Sync (single POST, parse JSON):gpt-image-1-generate, gpt-image-1-edit, openai-tts, openai-stt, openai-translate. Images come back base64 (data[0].b64_json) and are written to the run dir; audio (TTS) comes back as raw bytes; STT/Translate return text.
Stream (SSE):gpt-4o-chat (token deltas via choices.0.delta.content) and both gpt-image-2 nodes (partial-image frames + final). These set stream: true in the request body. There is no async-poll pattern anywhere in the OpenAI provider — unlike Meshy/Runway, nothing returns a task id to poll.
Status / error codes. Handlers treat any non-200 as a failure and surface OpenAI ... error <status>: <body>. Common: 400 bad params (e.g. an unsupported size for the chosen model), 401 bad/missing key, 429 rate limit / no quota, 403/organization_must_be_verified → org-verification (gpt-image-2 only, see below), 500 server. The gpt-image-2 handlers detect the org-verification body and rewrite it to a clear message pointing at the settings page.
Input-URI rules.
Images into edit/chat are resolved from local run-dir file paths. The edit nodes (gpt-image-1-edit, gpt-image-2-edit) read the upstream image file bytes and POST multipart — they do not accept a remote URL on the image port; the file must exist on disk. gpt-4o-chat's images port is more flexible: a value starting with http:///https:///data: is passed through as an image_url; anything else is treated as a local path, read, and base64-encoded into a data URI (png/jpg/jpeg/webp MIME inferred from extension).
Masks (gpt-image-1-edit, gpt-image-2-edit) are PNG with an alpha channel; the mask applies to the first image and marks where edits are allowed. The prompt must describe the whole desired image, not just the masked region.
Key gotchas (full list in "Capability boundaries"):
Translate is English-only and model-locked.openai-translate always runs whisper-1 under the hood (no model param); output is English text regardless of input language.
instructions (voice tone) only works on gpt-4o-mini-tts.tts-1/tts-1-hd silently ignore it.
STT temperature only affects whisper-1. The GPT-4o transcribe models ignore it; the handler only forwards non-zero values.
gpt-image-2 drops n>1, background, input_fidelity, and a user partial_images count — OpenAI rejects n>1 while streaming, and these params aren't valid/forwarded. Run the node multiple times for multiple images.
gpt-5.x are reasoning models: sampler params are dropped. On gpt-4o-chat, the handler omits temperature/top_p/frequency_penalty/presence_penalty for any gpt-5.* model (the API rejects them) and instead sends reasoning_effort (none/low/medium/high/xhigh). The sampler params are only forwarded — and only shown in the UI — for the legacy gpt-4o/4.1 models.
output_compression(int 0–100, default 90): only applied when output_format ≠ png.
moderation(enum, default auto): auto · low (less restrictive).
Handler always sets stream: true, partial_images: 0, and strips background/input_fidelity/n. For prompting craft + cost table, see the gpt-image-2 skill.
n(int 1–10, default 1): number of variations; only sent when >1.
size / quality / output_format / background: same enums/defaults as gpt-image-1-generate.
Inputs: image (req), prompt (req), mask (optional PNG). Image + mask are read from disk and uploaded multipart.
gpt-image-2-edit (stream, multipart)
Same param set as gpt-image-2-generate (size, quality, output_format, output_compression, moderation). No model param (fixed gpt-image-2), no n, no background.
Inputs: images (req, multiple, ≤10 — handler rejects >10), prompt (req), mask (optional, applies to first image).
language(enum, default auto): auto (omitted) · enesfrdeitptjakozharruhi.
response_format(enum, default text): text · json · verbose_json · srt · vtt. For json/verbose_json the handler extracts .text; for srt/vtt/text it returns the raw body. Output port is always Text — even SRT/VTT come back as a text string (save it as .srt/.vtt downstream).
temperature(float 0–1, step 0.1, default 0): only forwarded when non-zero; affects whisper-1 only.
prompt(string, optional): biasing hint (spellings, names, style).
openai-translate (sync, multipart)
No model param — fixed whisper-1, output is English text regardless of source language.
response_format(enum, default text): text · json · verbose_json · srt · vtt (same extraction logic as STT).
temperature(float 0–1, default 0): forwarded only when non-zero.
prompt(string, optional): must be in English.
gpt-4o-chat (stream) — display name "OpenAI Chat" (updated 2026-06-10)
reasoning_effort(enum, default medium): none · low · medium · high · xhigh. gpt-5.x models only — shown only when a gpt-5.* model is selected and sent only for those models.
response_format(enum, default text): text · json_object. json_object sends {"type":"json_object"} and enables JSON mode — JSON-mode only, no strict json_schema (see boundaries). Gotcha: OpenAI requires the literal word "json" to appear somewhere in the prompt (messages/system) when json_object is set, or the request errors. Instruct the model to "respond in JSON."
Inputs: messages (req, Text — becomes the user message text) and images (optional, multiple — each appended as an image_url content part for vision). A single user message is constructed; there is no multi-turn history or system-role port.
Recipes
All node ids below are real and in node_definitions.json.
Illustrated quote card (text → image).Text (your quote) → gpt-image-2-generate (size 1024x1536, quality high) → canvas. Iterate at quality low to save cost, then bump to high. Use gpt-image-1-generate (background transparent) for a cut-out sticker.
Narrated image caption (three OpenAI nodes, one key).Text (image brief) → gpt-image-1-generate → feed its image output into gpt-4o-chat's images port with a messages prompt like "Write a 2-sentence caption" → take the text output into openai-tts (voice coral) → an Audio clip of the spoken caption. (Chains text-gen → vision → audio-gen.)
Subtitle a voiceover.Audio → openai-stt (model gpt-4o-transcribe, response_format srt) → the Text output is SRT content; save as .srt. If the audio is in another language and you want English subtitles, swap openai-stt for openai-translate (same response_format srt).
Masked inpaint.Image + Mask + Text ("replace the sky with aurora") → gpt-image-2-edit (or gpt-image-1-edit for a single image + the n Count param). The prompt must describe the whole desired image; the mask marks where edits are allowed and applies to the first image. gpt-image-2-edit accepts up to 10 reference images for compositing.
In the nebula_nodes context
Node ids (8):gpt-image-1-generate, gpt-image-2-generate, gpt-image-1-edit, gpt-image-2-edit, openai-tts, openai-stt, openai-translate, gpt-4o-chat. Palette groups: the four image nodes under image-gen, the three audio nodes under audio-gen, chat under text-gen. (dalle-3-generate was removed 2026-06-10 — OpenAI shut down dall-e-2/3 on 2026-05-12.)
Ports. Image gen: in prompt (Text) → out image (Image). Image edit: in image/images+prompt+mask → out image. TTS: in text (Text) → out audio (Audio). STT/Translate: in audio (Audio) → out text (Text). Chat: in messages (Text) + images (Image, multiple) → out text (Text).
Chaining rules.
An Image output port wires straight into any image-input port (gpt-image-*-edit images, or gpt-4o-chatimages).
The edit nodes need the upstream image as a local file (they read bytes for multipart); a same-graph upstream image node satisfies this. gpt-4o-chat additionally accepts http(s)/data-URI strings on images.
Text outputs (chat, STT, translate) chain into any Text-input port — e.g. STT text → gpt-4o-chatmessages, or chat text → openai-ttstext.
There is no task-id port anywhere in this provider (no async-poll), so no "poll handle" chaining like Meshy/Runway.
How outputs render. Generated images write base64 to the run dir and display in the node's preview surface (DynamicNode/ModelNode); gpt-image-2 streams partial frames that swap in live as it renders. Chat text streams token-by-token into the node's text surface. Audio (TTS) renders as an inline <audio> player with a download control on the node. STT/Translate text (including SRT/VTT) renders as text.
Capability boundaries (what the OpenAI API supports but Nebula does NOT expose)
Do not promise these — they aren't wired. (From the audit guide's gap table, docs/api-guides/openai.md.)
Image variations (/v1/images/variations, dall-e-2) — no node; largely moot since OpenAI retired dall-e-2/3 on 2026-05-12.
background: transparent on gpt-image-2 — exposed only on GPT Image 1 / 1 Edit; the API doesn't support it on v2. For a transparent v2 result, chain a downstream remove-background node instead.
n>1 on the gpt-image-2 nodes — dropped (OpenAI rejects n>1 while streaming); only gpt-image-1-edit exposes Count (1–10). Run the node again for more images.
User-controllable streaming partial_images count — gpt-image-2 streams internally but the count is fixed at 0 (not a param).
input_fidelity on edits — not forwarded; gpt-image-2 always processes inputs at high fidelity.
Streaming TTS (stream_format: sse on gpt-4o-mini-tts) — the node writes the whole file; no live audio streaming.
Transcription timestamp_granularities (word/segment) — verbose_json is selectable but per-word/segment timestamps aren't surfaced.
Streaming / diarized transcription (gpt-4o-transcribe-diarize) — no streaming STT; the diarize model isn't in the enum.
Chat tool / function calling — no tools port.
Strict structured outputs (json_schema) — only json_object mode; no schema enforcement.
Audio-in/out chat (gpt-4o-audio-preview) — not wired.
Realtime API (gpt-realtime, speech-to-speech) — entire family unsupported.
Responses API (/v1/responses) — Nebula uses the older Chat Completions endpoint.
Embeddings / Moderations / Batch / Files / Assistants / Sora video — out of scope (Sora's Videos API + sora-2/sora-2-pro shut down 2026-09-24 anyway).
Coverage: roughly 60% of OpenAI's media-generation surface (images + audio + chat) is exposed; ~30% of the entire OpenAI platform.
