ComfyUI-NanoBanana2
A ComfyUI node using Google Gemini for prompt refinement and high-fidelity multi-reference image generation.
Nodes (30)
The one node you wire into everything else first
Turn any ComfyUI audio into a timestamped transcript
Know what a batch run will cost before you run it
See how big your prompt is before you pay for it
Turn text into a vector, for the RAG-ish workflows in your graph
Persist an embedding to disk so you don't re-pay for it
Upload a big file once, reuse the handle instead of re-sending it
Tell Google what to change, not how to change it
Google's Nano Banana, without the browser tab
Imagen 4 in ComfyUI, using the other Google image endpoint
Fill the masked hole with anything Gemini can imagine
Find out what your API key can actually access
Pick a Gemini model from one dropdown, with a sanity check built in
A chat node that remembers what it said last turn
Prompt a song out of Google's Lyria 3
The node that makes your Gemini calls look like they come from the US
Extend the canvas outward without breaking the illusion
Feed it a rough idea, get a prompt worth spending credits on
Turn Gemini's content filters up or down, per category
Force Gemini to return valid JSON, not a rambling answer
Gemini text generation with every knob exposed — no GPU required
Let Gemini write the Python and run it too
TextGen that can actually look things up — with citations
The tiny node that turns Gemini's thinking mode into data
Gemini-grade TTS that drops straight into ComfyUI's AUDIO type
Scripted two-voice dialogue, each speaker with their own voice
Veo 3 on your canvas — including the awkward wait
Ask Gemini what's actually in your image — up to four at once
OCR that actually reads small text — because JPEG smears it
Upload the PDF once, then ask it questions forever
ComfyUI-NanoBanana2
Yes, the name is ridiculous. No, we're not changing it.
"NanoBanana" is a community nickname for Google's Gemini image generation models. This started as a humble 3-node Gemini image generator. It has since gotten completely out of hand. We are now at 29 nodes covering text, vision, image generation (two different endpoints), audio, music, video, embeddings, file upload + reuse, Google Search grounding, code execution, multi-speaker TTS, audio transcription, OCR, and cost estimation. At some point this stopped being a ComfyUI node pack and became a full Gemini SDK replacement in node-graph form.
The name still fits, somehow.
Also available as part of ComfyUI-API-Toolkit alongside other API integrations.
v2.5 — audit release. Two parallel full-codebase reviews + 64 tests fixed 10 verified bugs, hardened secret handling / URL injection / SSRF / cache memory growth, and added 9 new nodes. See CHANGELOG.md.
Installation
ComfyUI Manager (recommended)
Search for NanoBanana2 in the ComfyUI Manager and install.
Registry
comfy node registry-install nanobanana2
Manual
git clone https://github.com/IxMxAMAR/ComfyUI-NanoBanana2
pip install google-genai
Getting an API Key
Go to aistudio.google.com, hit "Get API Key", copy it. That's it.
Paste it into the API Key node (password-masked) or set the GEMINI_API_KEY environment variable and the node will pick it up automatically.
Nodes
Config (6 nodes)
The connective tissue. Wire these into your generative nodes as needed.
| Node | What it does |
|---|---|
| API Key | Password-masked key input. Reads GEMINI_API_KEY env var if left empty. |
| Model Selector | Pick from text, image, or all models. Includes a custom override field for whatever Google released last Tuesday. |
| Safety Settings | Per-category harm thresholds. For when the defaults are either too strict or not strict enough for your workflow. |
| Thinking Config | Set thinking level and token budget. Defaults to NONE -- thinking is opt-in because it costs more and you probably don't need it for a caption node. |
| List Available Models | Queries your API key and returns what's actually accessible on your account. Useful when you're not sure if you have access to a preview model. |
| Token Counter | Count tokens for a given prompt before you burn them. Feed it your text and optional images, get back a number. |
Text (4 nodes)
| Node | What it does | |---|---| | Text Generation | Full-parameter text gen: temperature, top_p, top_k, thinking, seed. 33 model options including latest aliases, Gemini 3 previews, 2.5/2.0 stable, Gemma, and specialized models. If it's a knob, it's exposed. | | Prompt Refiner | Feed it a rough prompt, get back a polished one. Useful before hitting your image nodes. | | Multi-Turn Chat | Stateful conversation node. Maintains message history across runs. | | Structured Output | JSON schema-constrained generation. Tell it exactly what shape of data you want back. |
Image (6 nodes)
Two different endpoints, both covered.
| Node | What it does |
|---|---|
| Vision Analysis | Describe, analyze, or interrogate an image. Good for captioning or feeding into a downstream prompt. |
| Image Generation | Generate images via generate_content. Supports Nano Banana, Nano Banana 2, and Nano Banana Pro. Up to 4 reference images, full aspect ratio selection, seed control. |
| Imagen Image Generation | Dedicated Imagen 4 node using the generate_images endpoint. Ultra, Standard, and Fast variants. Different endpoint, different characteristics, same node graph. |
| Image Edit | Text-guided image editing. Describe what you want changed. |
| Inpaint | Mask-based inpainting. Feed it an image and a mask, tell it what should be there. |
| Outpaint | Extend an image outward. Choose your expansion direction. |
Audio (2 nodes)
| Node | What it does | |---|---| | Text-to-Speech | Convert text to speech using Gemini TTS models. 30+ prebuilt voices: Zephyr, Puck, Kore, Charon, and more. Flash and Pro variants. | | Music Generation | Generate music clips via Lyria 3. Clip and Pro model variants. Text prompt in, audio out. |
Video (1 node)
| Node | What it does |
|---|---|
| Video Generation | Text-to-video and image-to-video via Veo. Supports Veo 3.1 (including fast and lite), Veo 3.0, and Veo 2.0. Uses predictLongRunning under the hood because video takes a minute. |
Embeddings (2 nodes)
| Node | What it does |
|---|---|
| Text Embeddings | Generate text embeddings at 768 to 3072 dimensions. Includes task-type optimization for retrieval, clustering, classification, and semantic similarity use cases. |
| Save Embedding (.npy) | Write the embedding vector to disk as a NumPy .npy for downstream vector DBs / similarity search. Basename-sanitized. |
Files (2 nodes) — new in v2.1
| Node | What it does | |---|---| | Files Upload | Upload a local PDF / video / audio / image to the Gemini Files API. Returns a reusable URI (lives ~48h server-side) so you don't have to re-encode big files into every prompt. | | Ask Uploaded File | Pair with Files Upload. Send the URI + mime_type + a question, get text back. Supports PDF up to 1000 pages, video / audio up to 1 GiB. |
Tool-using Text (2 nodes) — new in v2.1
| Node | What it does |
|---|---|
| Text Gen + Google Search | TextGen with the GoogleSearch grounding tool wired in. Returns the answer plus a citations_json list of (url, title) sources the model used. For current events, recent product info, etc. |
| Text Gen + Code Execution | TextGen with the ToolCodeExecution tool. Returns three strings: the final answer, any Python the model ran, and the execution output. Math, stats, plotting, JSON wrangling. |
Vision + Audio additions (2 nodes) — new in v2.1
| Node | What it does |
|---|---|
| Vision OCR (lossless PNG) | OCR-tuned Vision node. Three modes: plain_text, structured_json (returns line-level bboxes), or markdown. Always-PNG encoding so small text isn't smeared by JPEG. |
| TTS Multi-Speaker Dialogue | Proper two-voice TTS using MultiSpeakerVoiceConfig. Write Alice: Hi Bob.\nBob: Hi Alice. and assign each speaker a voice. |
| Audio Transcribe | Feed a ComfyUI AUDIO, get a transcript back. Optional [HH:MM:SS] timestamps. |
Utilities (new in v2.1)
| Node | What it does |
|---|---|
| Cost Estimator | Counts input tokens via the free count_tokens API, multiplies by a model price table + your output-token estimate + a run multiplier, returns total USD plus a breakdown string. Pre-flight expensive batched workloads. |
Supported Models
generateContent (text and multimodal)
Latest aliases
gemini-pro-latest,gemini-flash-latest,gemini-flash-lite-latest
Previews
gemini-3-pro-preview,gemini-3-flash-previewgemini-3.1-pro-preview,gemini-3.1-flash-lite-preview
Stable
gemini-2.5-pro,gemini-2.5-flash,gemini-2.5-flash-litegemini-2.0-flash,gemini-2.0-flash-lite,gemini-2.0-flash-001,gemini-2.0-flash-lite-001
Gemma
gemma-3-1b-it,gemma-3-4b-it,gemma-3-12b-it,gemma-3-27b-itgemma-3n-e2b-it,gemma-3n-e4b-itgemma-4-26b-a4b-it,gemma-4-31b-it
Specialized
gemini-robotics-er-1.5-preview,gemini-robotics-er-1.6-previewgemini-2.5-computer-use-previewdeep-research-pro-previewnano-banana-pro-preview
generateContent (image output)
gemini-3.1-flash-image-preview-- Nano Banana 2gemini-3-pro-image-preview-- Nano Banana Progemini-2.5-flash-image-- Nano Banana
predict (Imagen)
imagen-4.0-ultra-generate-001imagen-4.0-generate-001imagen-4.0-fast-generate-001
generateContent (TTS audio output)
gemini-2.5-flash-preview-ttsgemini-2.5-pro-preview-ttsgemini-3.1-flash-tts-preview- 30+ prebuilt voices
predict (Lyria music)
lyria-3-pro-previewlyria-3-clip-preview
predictLongRunning (Veo video)
veo-3.1-generate-preview,veo-3.1-generate-fast,veo-3.1-generate-liteveo-3.0-generate-001,veo-3.0-generate-fastveo-2.0-generate-001
embedContent (embeddings)
gemini-embedding-001gemini-embedding-2-preview
All nodes also have a custom_model override field. When Google drops something new you can use it immediately without waiting for a package update.
Aspect Ratios
1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9
Technical Notes
A few things that were done deliberately:
- IS_CHANGED on all nodes -- every node re-executes on every run even with identical inputs. Generative nodes should be generative.
- Retry with jittered exponential backoff -- transient API errors are retried with random jitter to avoid thundering-herd retries when multiple workers hit a 429 wave.
- Bounded, hashed client cache -- LRU cap of 16 clients, keyed by SHA-256 of the API key (the raw key never lives in a dict key). Rotating keys across workflows can't OOM the worker.
- Full chunk iteration -- responses scan all parts, not just
parts[0]. You won't silently lose content from multi-part responses. - Image candidates returned as a batch -- ImageGen with
candidate_count > 1returns ALL images as a batched IMAGE tensor (previously you paid for 4, got 1). - Mask auto-resize -- Inpaint / ImageEdit silently match mask dims to image dims (Gemini 400s on mismatch).
- Lossless PNG option -- Vision node has a
losslesstoggle for OCR / small-text / fine-grained tasks; new VisionOCR node uses it by default. - API-key redaction -- every error string surfaced to the UI / logs is run through a regex that strips
AIza[...]keys andx-goog-api-key:/Authorization:headers. - URL-path injection guarded -- custom Lyria / Veo model IDs are validated against
^[A-Za-z0-9._-]+$so a malicious../other_endpointcan't escape the/models/path. - SSRF guards -- Veo only downloads from
*.googleapis.com/*.googleusercontent.com;download_filedefaultsallow_redirects=Falseand enforces a 256 MiB cap by default. - Safety refusals are descriptive -- when Gemini refuses a request, the error tells you which category triggered it and surfaces the model's own explanation.
- Tooltips everywhere -- hover over any input for a description of what it does.
- Password-masked API keys -- the key input field is masked.
- Environment variable fallback -- set
GEMINI_API_KEYand all nodes pick it up automatically. Quote stripping handles the commonGEMINI_API_KEY="AIza...".envmistake.
Requirements
- Python 3.10+
google-genai >= 0.8.0- A Google AI Studio API key
Tests
python -m pytest tests/
64 tests, no network. Covers secret redaction, model-ID sanitization, mask/image conversions, retry control flow, response size caps, every new node's registration, and regressions on each verified v2.0 bug.
License
MIT
Made by IxMxAMAR