一键在 Manus 中运行任何 Skill

gemini

星标13

分支4

更新时间2026年6月10日 17:02

Google's Gemini / Imagen / Veo / Lyria models inside Nebula — text & vision chat, Imagen 4 text-to-image, Nano Banana image gen/edit/compose, Veo 3.1 video with sound, Lyria 3 music, Gemini TTS narration, text embeddings, and a Style Reference utility. Activate when the user configures any gemini-chat, imagen-4-generate, nano-banana, veo-3, lyria-3, gemini-tts, gemini-embeddings, or style-reference node, or asks about Gemini / Imagen / Veo / Nano Banana / Lyria in Nebula. For prompt craft also load the topic files (nano-banana.md, veo.md, imagen.md, gemini-text.md). Sourced from ai.google.dev (image-generation, imagen, video, speech-generation, music-generation, embeddings) and the Nebula audit guide docs/api-guides/google.md, cross-checked against backend/data/node_definitions.json + backend/handlers/google_gemini.py + backend/handlers/veo.py on 2026-06-04.

安装

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。

在 Manus 中运行

来源

JustinPerea

JustinPerea/nebula-nodes

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Gemini Skill

Nebula exposes Google's generative stack as 8 nodes, all keyed on one GOOGLE_API_KEY and all hitting generativelanguage.googleapis.com/v1beta directly (the Gemini Developer API, not Vertex). This file is the authoritative map of the real Nebula node IDs, param keys, ports, and execution patterns. For prompt-writing depth, load the topic files listed under Routing.

When to use

The user configures or asks about any Google node: gemini-chat, imagen-4-generate, nano-banana, veo-3, lyria-3, gemini-tts, gemini-embeddings, style-reference.
The user asks "which Gemini / Imagen / Nano Banana / Veo / Lyria model should I use?" or how to wire one into a graph.
Editing the handlers backend/handlers/google_gemini.py or backend/handlers/veo.py, or the Google node defs in backend/data/node_definitions.json.
The user describes an image/video/music/TTS/embedding task that a Google node could serve.

Universal rules

Auth. Every node sends the header x-goog-api-key: <GOOGLE_API_KEY> (not Authorization: Bearer). The env var is GOOGLE_API_KEY, read from .env at the repo root; restart the backend after adding it. One key covers all 8 nodes. veo-3 also declares FAL_KEY because it can route through FAL — but the direct Google path documented here needs only GOOGLE_API_KEY.
Base URL. https://generativelanguage.googleapis.com/v1beta/models. The endpoint verb is per-node: :generateContent (Nano Banana, Lyria, TTS, Style Reference), :streamGenerateContent?alt=sse (Gemini chat), :predict (Imagen), :embedContent (Embeddings), :predictLongRunning + operation polling (Veo).
Execution patterns (confirmed from handlers):
- stream (SSE) — gemini-chat only. Hits :streamGenerateContent?alt=sse; deltas read from candidates.0.content.parts.0.text; 60 s timeout.
- sync (one POST, blocks for the result) — nano-banana, imagen-4-generate, lyria-3, gemini-tts, gemini-embeddings, style-reference. Timeouts: 120 s images/TTS, 180 s Lyria, 60 s embeddings.
- async-poll — veo-3 only. POST to :predictLongRunning returns an operation name; the handler polls GET /v1beta/{name} every 3 s up to 300 times (~15 min ceiling), emits progress, then downloads the finished MP4 to the run dir. Expect minutes, not seconds.
Status / error codes. Handlers treat any non-200 as a hard failure and raise "<Service> API error <status>: <body>". Common: 400 bad params (e.g. sending both thinkingLevel and thinkingBudget to Gemini 3 → 400; an unsupported aspect/size for the chosen model), 401/403 bad or missing key, 429 rate/quota, 5xx transient. Veo additionally surfaces an error object inside the completed operation.
Input-URI rules. Image inputs on gemini-chat, nano-banana, lyria-3, and veo-3 accept three forms and the handler converts each: a local file path (base64-inlined; png/jpg/jpeg/webp mapped, anything else defaults to image/png), an http(s):// URL (sent as fileData/file_data; for Veo it's downloaded then inlined), or a data: URI (split into mime + base64). So you can wire another node's Image output straight in. style-reference is the exception — it takes no input port; you set its filePath param to a local image.
Key gotchas.
- Gemini 3 vs 2.5 thinking is mutually exclusive. thinkingLevel only shows on the Gemini 3 models; thinkingBudget only on the 2.5 models. The handler sends one or the other — never both (the API 400s on both).
- Model IDs are the literal enum strings. The exact strings below (e.g. gemini-3.1-flash-image-preview, gemini-3-flash-preview, lyria-3-clip-preview, veo-3.1-generate-preview) are the values the node's model enum actually sends. Do not "modernize" them by stripping or adding -preview — most are pinned preview ids, but a few are stable ids with no suffix (gemini-3.5-flash, gemini-3.1-flash-lite, the 2.5 text models). Use each id exactly as listed. (The prompting-craft topic files sometimes use shorthand like imagen-4; the IDs in this file are the ground truth for wiring.)
- Nano Banana / Imagen / Veo have no transparent-background mode. Ask for "white background" in the prompt for stickers/cutouts.
- Veo outputs expire on Google's side (~2 days). Nebula downloads the MP4 to the run folder automatically on completion, so the local copy is yours; don't rely on the remote URI later.
- Lyria WAV uses a proto-enum mime. When outputFormat=wav on the Pro model, the handler sends responseFormat.audio.mimeType = "AUDIO_WAV" (not the literal "audio/wav", which the API rejects). WAV is Pro-only; the Clip model returns MP3.
- TTS audio is raw PCM. The model returns 24 kHz / 16-bit / mono PCM; the handler wraps it into a .wav. Output is always WAV regardless of model.
- Embeddings output is text, not a vector type. The embedding port is a JSON-stringified float array (Text), and dimensions is the count as a string — so downstream nodes receive text you parse yourself.

Pick the right node

One row per Nebula Google node. IDs, ports, and param keys are ground-truth from node_definitions.json.

Node (display)	Node ID	Category · pattern	Endpoint verb	Key inputs → outputs	Model enum (literal values)	Headline params
Gemini	`gemini-chat`	text-gen · stream	`:streamGenerateContent?alt=sse`	`messages` (Text, req), `images` (Image, multi) → `text`	`gemini-3.5-flash` (default, stable), `gemini-3.1-pro-preview`, `gemini-3-flash-preview`, `gemini-3.1-flash-lite` (stable), `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`	`temperature`, `max_tokens`, `system`, `thinkingLevel` (G3) / `thinkingBudget` (2.5), `top_p`, `top_k`, `stop_sequences`, `response_format`
Imagen 4	`imagen-4-generate`	image-gen · sync	`:predict`	`prompt` (Text, req) → `image`	`imagen-4.0-generate-001` (default), `imagen-4.0-ultra-generate-001`, `imagen-4.0-fast-generate-001`	`aspectRatio`, `numberOfImages`, `seed`, `enhancePrompt`, `imageSize` (1K/2K, base+Ultra only), `personGeneration`
Nano Banana	`nano-banana`	image-gen · sync	`:generateContent`	`prompt` (Text, req), `images` (Image, multi) → `image`, `text`	`gemini-3.1-flash-image-preview` (default, "Nano Banana 2"), `gemini-3-pro-image-preview` ("Pro"), `gemini-2.5-flash-image` ("2.5 Flash")	`aspect_ratio`, `imageSize` (512/1K/2K/4K)
Veo 3.1	`veo-3`	video-gen · async-poll	`:predictLongRunning` → poll	`prompt` (Text, req), `image` (First Frame), `last_frame` (Last Frame), `video` (Extend Video) → `video`, `source_uri`	`veo-3.1-generate-preview` (default), `veo-3.1-fast-generate-preview`, `veo-3.1-lite-generate-preview`, `veo-3.0-generate-001`, `veo-3.0-fast-generate-001`, `veo-2.0-generate-001`	`aspectRatio`, `duration`, `resolution`, `personGeneration`, `seed` (direct); `negative_prompt`/`safety_tolerance` (FAL path only)
Lyria 3	`lyria-3`	audio-gen · sync	`:generateContent`	`prompt` (Text, req), `images` (Image, multi) → `audio`, `text` (Lyrics)	`lyria-3-clip-preview` (default, 30 s), `lyria-3-pro-preview` (full song)	`outputFormat` (MP3/WAV — WAV is Pro-only)
Gemini TTS	`gemini-tts`	audio-gen · sync	`:generateContent`	`text` (Text, req) → `audio`	`gemini-3.1-flash-tts-preview`, `gemini-2.5-flash-preview-tts` (default), `gemini-2.5-pro-preview-tts`	`voiceName` (30 voices)
Gemini Embeddings	`gemini-embeddings`	utility · sync	`:embedContent`	`text` (Text, req) → `embedding`, `dimensions`	`gemini-embedding-001` (default, text), `gemini-embedding-2-preview` (multimodal — reaches API as text only)	`taskType` (8 enums), `outputDimensionality` (768/1536/3072)
Style Reference	`style-reference`	utility · sync	`gemini-2.5-flash:generateContent`	no input port → `image` (Reference), `style_description` (Style)	fixed `gemini-2.5-flash`	`filePath` (req), `mode`, `manual_description`, `focus`, `strength`

Param reference

Defaults, ranges, and enums per node, straight from node_definitions.json + handler behavior.

`gemini-chat`

model (enum, req, default gemini-3.5-flash) — seven values; see table (updated 2026-06-10: gemini-3.5-flash is the new stable default, replacing gemini-2.5-flash; gemini-3.1-flash-lite replaced the old gemini-3.1-flash-lite-preview id). The display picks the family.
max_tokens (int, default 8192, 1–65535) → generationConfig.maxOutputTokens.
temperature (float, default 1, 0–2, step 0.1). Keep at 1.0 for Gemini 3 (lowering can loop/degrade).
system (textarea) → systemInstruction.
thinkingLevel (enum, Gemini 3 models only: ""/minimal/low/medium/high) → thinkingConfig.thinkingLevel. "" = model default.
thinkingBudget (int, 2.5 models only, 0–65536) → thinkingConfig.thinkingBudget. Mutually exclusive with thinkingLevel.
top_p (float, 0–1, step 0.05), top_k (int, hint 64), stop_sequences (comma-separated string).
response_format (enum, default text/plain; or application/json). JSON mode is now wired (2026-06-05): application/json sets generationConfig.responseMimeType = "application/json" so the reply is valid JSON. There is still no responseSchema field, so a strict schema must be prompt-engineered. Unit-tested.
Output: single text port. Streams token-by-token to the canvas.

`imagen-4-generate`

model (enum, req, default imagen-4.0-generate-001) — base / ...ultra... / ...fast....
aspectRatio (enum, default 1:1): 1:1, 4:3, 3:4, 16:9, 9:16.
numberOfImages (int, default 1, 1–4) → sampleCount. Note: only the first prediction is saved to the image port.
seed (int, optional) — reproducibility.
enhancePrompt (bool, default false) — Google rewrites the prompt for richer detail.
imageSize (enum, default 1K; 1K/2K) — visible only on base + Ultra (not Fast).
personGeneration (enum, default allow_adult): allow_all / allow_adult / dont_allow.

`nano-banana`

model (enum, default gemini-3.1-flash-image-preview). Pro = gemini-3-pro-image-preview; original = gemini-2.5-flash-image.
aspect_ratio (enum, default 1:1): 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 21:9. Extreme ratios 1:4, 4:1, 1:8, 8:1 appear only on Nano Banana 2 (...3.1-flash-image-preview).
imageSize (enum, default 1K; visible on Nano Banana 2 + Pro, not 2.5 Flash): 1K, 2K, 4K, plus 512 on Nano Banana 2 only.
Inputs: up to N reference images on images (multi). Outputs both an image and any model text.
Sends responseModalities: ["IMAGE","TEXT"]; aspect/size go into generationConfig.imageConfig.

`veo-3`

Params split three ways in the def (this is a dual-provider node):

sharedParams (both Google + FAL): aspectRatio (default 16:9; 16:9/9:16), duration (default 8; valid set varies by model — 4/6/8 on Veo 3.x & Veo 3, 5/6/8 on Veo 2), resolution (default 720p; 720p/1080p, plus 4k on full Veo 3.x and Veo 3 — not Lite), personGeneration (allow_all/allow_adult/dont_allow).
directParams (Google path): model (default veo-3.1-generate-preview, six values), seed (int; not on Veo 2).
falParams (FAL path only): negative_prompt (string), seed (int), safety_tolerance (enum 1–6, default 4). The direct Google path does not expose a negative prompt.
Handler maps duration → durationSeconds (int), strips a trailing s. First/last frame come from the image / last_frame ports, not params.
Video extension (Google path, 2026-06-08). A 4th input port video ("Extend Video") continues an existing clip; a 2nd output source_uri carries this generation's Veo files/... URI. To extend in-graph, wire an upstream veo-3 node's source_uri → a downstream veo-3 node's video input. Veo only accepts its own generated source URI (re-uploaded/external video → 400 "must be a video generated by VEO"), so wiring the regular video (downloaded file) output into Extend Video fails by design — use source_uri. Extension is veo-3.0/3.1 generate+fast only (not lite, not veo-2.0), output is locked to 720p, and the source URI is valid only ~2 days. Cannot combine with image/last_frame.
Reference images ("ingredients to video") are NOT available. Documented for Veo 3.1, but the live Gemini API rejects them ("Your use case is currently not supported") on both veo-3.1-generate-preview and veo-3.1-fast-generate-preview with an AI-Studio key (verified 2026-06-08). There is no reference-image port; don't promise subject/wardrobe consistency via Veo refs.

`lyria-3`

model (enum, req, default lyria-3-clip-preview): Clip (≤30 s) or lyria-3-pro-preview (full song).
outputFormat (enum, default mp3; visible only on Pro): mp3 / wav. WAV → proto enum AUDIO_WAV over the wire.
Inputs: prompt (req) + optional images (mood/scene reference). Outputs audio + text (Lyrics).
Sends responseModalities: ["AUDIO","TEXT"].

`gemini-tts`

model (enum, default gemini-2.5-flash-preview-tts): gemini-3.1-flash-tts-preview (added 2026-06-10) · gemini-2.5-flash-preview-tts · gemini-2.5-pro-preview-tts.
voiceName (enum, default Kore) — 30 prebuilt voices, each labeled with a character: e.g. Kore (Firm), Puck (Upbeat), Zephyr (Bright), Charon (Informative), Fenrir (Excitable), Leda (Youthful), Aoede (Breezy), Enceladus (Breathy), Algieba (Smooth), Sulafat (Warm), Achird (Friendly), Gacrux (Mature) … through Sadaltager (Knowledgeable). Pick by the vibe label.
Input: text (req) — this is the script to speak. Output: audio (WAV).
Single-speaker only; sends speechConfig.voiceConfig.prebuiltVoiceConfig.voiceName.

`gemini-embeddings`

model (enum, default gemini-embedding-001) or gemini-embedding-2-preview (multimodal model, but the node only feeds it text).
taskType (enum, default SEMANTIC_SIMILARITY): SEMANTIC_SIMILARITY, RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT, CLASSIFICATION, CLUSTERING, CODE_RETRIEVAL_QUERY, QUESTION_ANSWERING, FACT_VERIFICATION. Match query vs document: embed stored docs with RETRIEVAL_DOCUMENT, embed the live search string with RETRIEVAL_QUERY — mismatching them hurts recall.
outputDimensionality (enum, default 768): 768 / 1536 / 3072. Smaller = cheaper storage & faster cosine; 3072 = max fidelity.
Input: text (req). Outputs: embedding (JSON array as Text) + dimensions (count as Text).

`style-reference`

A local utility that turns one image into a reusable look. No input port — set filePath directly.

filePath (file, req) — the reference image on disk.
mode (enum, default auto): auto (Gemini 2.5 Flash writes the description), manual (you supply it), passthrough ("Image only" — emit just the Reference image, no text).
manual_description (textarea) — visible only when mode=manual; the style text you author.
focus (enum, default all; visible only when mode=auto): all (palette + lighting + medium + mood), palette, lighting, medium (texture). Narrow it to transfer just one axis.
strength (float, default 0.7, 0–1, step 0.05) — how strongly the description leans into the style.
Outputs: image (Reference, the original image re-emitted) + style_description (Style, Text). Wire Style → a generator's prompt and Reference → that generator's images input.

Recipes

Concrete graphs using real node IDs.

Illustrated short with a soundtrack. imagen-4-generate (hero still, aspectRatio=16:9) → wire its image into veo-3's First Frame (image port), prompt the motion (model=veo-3.1-generate-preview, resolution=1080p). In parallel, lyria-3 (model=lyria-3-clip-preview, prompt the mood) produces a 30 s bed. Mix the Lyria audio under the Veo video downstream. Remember Veo is async-poll — it'll take minutes.
Style-locked product shots. style-reference (set filePath to a reference photo, mode=auto, focus=all, strength=0.7) → route style_description (Style) into nano-banana's prompt and the Reference image into its images input, then append your product description. Use model=gemini-3-pro-image-preview for legible packaging text; bump imageSize=2K. Every render inherits the reference's look and stays product-consistent across angles.
Narrated explainer with slides. gemini-chat (system="You are a concise narrator", response_format=text/plain) writes a script from a topic → feed its text into gemini-tts (voiceName=Charon for an informative read) for spoken VO. Generate matching imagen-4-generate stills for each slide. (Keep narration single-voice — TTS here is single-speaker.)
Searchable knowledge base (RAG backbone). For each document: gemini-embeddings (taskType=RETRIEVAL_DOCUMENT, outputDimensionality=768) → store the embedding text (parse the JSON array). At query time, embed the user's question with taskType=RETRIEVAL_QUERY and cosine-rank. Nebula builds the index yourself — there's no managed File Search node.

In the nebula_nodes context

Node IDs (8): gemini-chat, imagen-4-generate, nano-banana, veo-3, lyria-3, gemini-tts, gemini-embeddings, style-reference.
Handler files: backend/handlers/google_gemini.py (chat, nano-banana, imagen4, lyria3, tts, embeddings — and Style Reference's gemini-2.5-flash describe call) and backend/handlers/veo.py (Veo submit + poll + download). Param-build logic lives there; consult it before changing a param name.
Ports & chaining.
- Image outputs (imagen-4-generate.image, nano-banana.image, style-reference.image) plug into any node's Image-typed input (nano-banana.images, veo-3.image/last_frame, gemini-chat.images, lyria-3.images).
- gemini-chat.text → any Text input (e.g. gemini-tts.text, nano-banana.prompt, imagen-4-generate.prompt).
- style-reference.style_description (Text) → a generator's prompt; pair its image (Reference) into that generator's images.
- gemini-embeddings emits Text (a JSON array + a count), not a numeric vector type — parse it downstream.
How outputs render. Images/video/audio are saved to the run dir and served via the outputs route; the canvas previews them (image/video/audio players). gemini-chat streams text live into its node.
Where they appear in the palette: text-gen (Gemini), image-gen (Imagen 4, Nano Banana), video-gen (Veo 3.1), audio-gen (Lyria 3, Gemini TTS), utility (Gemini Embeddings, Style Reference).
Veo dual-provider: veo-3 carries both GOOGLE_API_KEY and FAL_KEY. The direct Google path (this skill) uses directParams; switching to FAL swaps in falParams (negative_prompt, safety_tolerance) — the negative prompt is only reachable via FAL.

Capability boundaries (what the API can do that Nebula does NOT expose)

Never promise these through the current nodes — they exist in the Google API but are not surfaced. (Source: gap table in docs/api-guides/google.md.)

Gemini text tools — all absent. No function calling, no Google Search grounding, no URL-context tool, no code execution. gemini-chat is plain generateContent with vision-in.
Structured output: JSON mode works, no schema. response_format=application/json is honored (sets generationConfig.responseMimeType, wired 2026-06-05), but there's no responseSchema — strict schemas must be prompt-engineered.
Vision-in only for understanding. gemini-chat accepts images; video / audio / PDF understanding is not wired.
Nano Banana: image-search grounding and explicit thinkingLevel control are not surfaced (text-to-image, edit, and multi-image compose are).
Imagen edit / inpaint / customize / upscale: not available — those live in Vertex AI, not the Gemini Developer API Nebula targets. (Not a Nebula gap, an API-surface boundary.)
Veo: reference images (up to 3) and video extension (longer-than-8 s clips, up to ~148 s) are not exposed on the direct path; negative prompt only via the FAL route.
TTS: multi-speaker (2-voice) dialogue is not surfaced — single-speaker only (the 3.1 Flash TTS preview model joined the enum 2026-06-10).
Lyria: Lyria RealTime (streaming/WebSocket music, weighted-prompt steering) is not exposed — only the batch Clip/Pro models.
Embeddings: true multimodal embedding (image/audio/video/PDF via Embedding 2) only reaches the API as text here.
Operational endpoints — none: no Live API (realtime voice), no Computer Use, no File Search / managed RAG, no Batch API, no context caching, no token counting.
Safety: only personGeneration (Imagen/Veo) is adjustable; there is no general safetySettings block.

Coverage is roughly the 8 core generative modalities (text, Imagen, Nano Banana, Veo, Lyria, TTS, embeddings, + a style-extraction helper); the gaps are advanced text-model tooling and operational/streaming endpoints, not missing media types.

Routing

Prompt-writing depth lives in topic files — load the relevant one before authoring the text that feeds a node:

nano-banana.md — Gemini image family: narrative prompting, editing, multi-image composition, in-image text.
veo.md — Veo 3.x: prompt structure, camera language, audio, refs, extension, first/last frame.
imagen.md — Imagen 4 text-to-image (uses imagen-4 shorthand; wire with the real imagen-4.0-* IDs above).
gemini-text.md — Gemini text prompting, structured output, thinking mode.
reference/model-ids.md — full model-ID catalog incl. Live, computer-use, deep-research (broader than the 8 Nebula nodes).
reference/official-docs.md — canonical ai.google.dev URLs to fetch when content smells stale.

These topic files are prompt-craft references synthesized 2026-04-16 and may lag the live API; the node IDs/params in this SKILL.md were re-verified against the repo on 2026-06-04 (model enums updated 2026-06-10) and are authoritative for wiring.

Sources

Gemini API overview — https://ai.google.dev/gemini-api/docs
Image generation & editing (Nano Banana) — https://ai.google.dev/gemini-api/docs/image-generation
Imagen — https://ai.google.dev/gemini-api/docs/imagen
Video generation (Veo) — https://ai.google.dev/gemini-api/docs/video
Speech generation (TTS) — https://ai.google.dev/gemini-api/docs/speech-generation
Music generation (Lyria) — https://ai.google.dev/gemini-api/docs/music-generation
Embeddings — https://ai.google.dev/gemini-api/docs/embeddings
Model catalog — https://ai.google.dev/gemini-api/docs/models
Nebula audit guide (node table, params, gap table) — docs/api-guides/google.md

gemini

同仓库更多 Skills

同仓库更多 Skills

Gemini Skill

When to use

Universal rules

Pick the right node

Param reference

gemini-chat

imagen-4-generate

nano-banana

veo-3

lyria-3

gemini-tts

gemini-embeddings

style-reference

Recipes

In the nebula_nodes context

Capability boundaries (what the API can do that Nebula does NOT expose)

Routing

Sources

Gemini Skill

When to use

Universal rules

Pick the right node

Param reference

gemini-chat

imagen-4-generate

nano-banana

veo-3

lyria-3

gemini-tts

gemini-embeddings

style-reference

Recipes

In the nebula_nodes context

Capability boundaries (what the API can do that Nebula does NOT expose)

Routing

Sources

`gemini-chat`

`imagen-4-generate`

`nano-banana`

`veo-3`

`lyria-3`

`gemini-tts`

`gemini-embeddings`

`style-reference`

`gemini-chat`

`imagen-4-generate`

`nano-banana`

`veo-3`

`lyria-3`

`gemini-tts`

`gemini-embeddings`

`style-reference`