Nanaix_Image

The name is a lie, sort of. Nanaix_Image doesn't run anything on your machine - it's an API wrapper that sends your prompt and up to eight reference images to ai.nanaix.com and drops a finished image back into your graph as a native ComfyUI IMAGE tensor. No model download, no VRAM, no sampler settings. You trade all of that for the one thing ComfyUI can't give you locally: a closed model. This node is the image-edit half of nan-boop's NanAix_comfyui pack, and it's for when you specifically want gpt-image-2 or nano-banana-2 / nano-banana-pro doing reference-based editing.

Why reach for it instead of the official ComfyUI API Nodes? Cost, mostly, plus region and billing. NanAix is one of those single-key reseller/proxy services that sit in front of the OpenAI and Google image APIs - think kie.ai or cometapi, but smaller and quieter. The README is bilingual (English + 中文) and its install examples point at a Chinese portable ComfyUI build (aki-v3), which tells you who it's really aimed at: if the official OpenAI/Google APIs are a wall where you live, a reseller key is often the only door. Honest caveat: this service has basically zero Reddit footprint as of mid-2026. You're trusting a small proxy, so you're not buying community consensus - you're buying access.

How it works

The node routes by model-name prefix. Type gpt-image-2 and it hits the image-2 Images API's edit endpoint with your reference images serialized in; type nano-banana-2 and it switches to the async task flow - POST /api/generate, then poll GET /api/result?id=task-id until a success or failure status arrives. Both paths come back as one or more decoded images, batched into a standard IMAGE tensor.

Before upload, the node normalizes your references: big images get downscaled to a 1536px max edge and JPEG-compressed at quality 85 to keep the request from timing out. A batched input on any slot is expanded into individual images. There's no mask or inpainting input in this MVP, so local region edits aren't a thing here - it's whole-image editing or nothing.

The inputs that matter

• image_1 - the main reference. Subject, composition, layout. At least one image slot must be connected or the node refuses to run.
• image_2 through image_8 - optional cues for style, lighting, palette, environment. The author's suggested pattern: subject in slot 1, style in slot 2, extras beyond.
• prompt, model, api_key - the free-text trio. model is deliberately free-form so future model names work without a node update; unsupported prefixes are rejected before any network call.
• resolution_preset / width / height - presets like square, landscape_hd, up to square_4k, or custom to keep the raw fields.
• n - image count. Above 1 you get a batch that still plugs straight into PreviewImage / SaveImage.

Everything else - background (set transparent for cutout PNGs on gpt-image-2), style (vivid), moderation, output_compression, partial_images, stream - is gpt-image-2-only and quietly ignored by banana models. official_website is a display-only link, never sent.

Install

Easiest: ComfyUI Manager → search "Nanaix", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/nan-boop/NanAix_comfyui.git nanaix_Comfy
cd nanaix_Comfy
pip install -r requirements.txt

Restart ComfyUI, then find Nanaix_Image under the "Nanaix" category. The dependency list is just Pillow and numpy - refreshingly light for a pack that generates images.

Common issues

• Invalid API key - wrong, missing, or pasted somewhere it isn't being read. It must go in the node's visible api_key field.
• Images API is not supported for this platform - your key's account group lacks image access. You need a key from an image-enabled group.
• No available compatible accounts - upstream is out of slots right now. Retry later.
• Timeouts - the classic with big edits. Fewer reference images, smaller references, or a smaller output. Banana jobs that overrun report the task id so you can re-query it manually.
• Config gotcha - after a successful run the node saves your settings to nanaix_config.json in the pack folder and prefills new nodes, but a blank visible key still errors even if an old key is in the file. And since that file holds a credential: don't commit it, don't paste your key into shared workflow JSON. This is an API node - it ships your key and calls the network by design, which is exactly the shape of thing that's been weaponized in this ecosystem before. This one's a small, readable, open codebase (that's reassuring), but read a fresh pack before you trust it, always.