Nanaix_Text

No GPU, no model folder, no UNET drama. Nanaix_Text takes a prompt, calls ai.nanaix.com, and hands you back a native ComfyUI IMAGE tensor you can wire straight into PreviewImage, SaveImage, an upscaler, or anything else downstream. It's the text-to-image half of the NanAix_comfyui pack by nan-boop, and it exists for one reason: the models it generates with - gpt-image-2 and nano-banana-2 / nano-banana-pro - are closed. You cannot download them, so an API node is the only door into your workflow. Think of it as a much cheaper, single-key stand-in for the official ComfyUI API Nodes if you're outside the regions those bill nicely in.

Here's the honest picture before you get excited. NanAix is a reseller/proxy sitting in front of OpenAI's and Google's image APIs, and as of mid-2026 it has essentially zero Reddit footprint - a loose search turns up nothing real, which is the opposite of the community buzz around, say, ComfyUI's own Partner Nodes. The README is bilingual and aimed partly at a Chinese-market ComfyUI build, which is the giveaway: this is the kind of service people use when the first-party API is blocked, unpayable, or just overpriced where they live. You're trading community consensus for access. Fine - just know what you're buying.

How it works

The node routes by model prefix. gpt-image-* goes to the image-2 Images API's text-to-image endpoint with a deliberately conservative payload (model, prompt, size, quality, format, b64_json) - the author keeps the request lean because unstable optional fields have caused upstream failures before. nano-banana-* takes the async path instead: submit with POST /api/generate, then poll GET /api/result?id=task-id until the task finishes. Either way, the returned images are decoded and batched into a standard IMAGE tensor, so it behaves like a local generator node on the canvas even though the pixels were born in a datacenter.

The inputs that matter

• prompt - the whole job. Multiline, with dynamic prompts enabled. A specific prompt gets you a controllable result.
• model - free-form text, defaulting to gpt-image-2. Type nano-banana-2 or nano-banana-pro to switch routes. Anything not starting with gpt-image- or nano-banana- is rejected before any request is sent.
• api_key - one key for both model families. It must be filled in the visible field; saved config prefills new nodes but a blank field still errors.
• resolution_preset / width / height - presets from square up to square_4k, or custom to keep manual control.
• quality and output_format - high/medium/low, and png/webp/jpeg.

The rest (background - set transparent for PNG cutouts - style, moderation, output_compression, partial_images, stream) are gpt-image-2 controls that banana models ignore. If you don't need transparent backgrounds or SSE streaming, nano-banana-* is the simpler, more predictable path.

Install

ComfyUI Manager → search "Nanaix" and install, or do it by hand:

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 and search Nanaix_Text (it lives under the "Nanaix" category). Dependencies are just Pillow and numpy - no torch, no checkpoint downloads, because there are no local weights to fetch.

Common issues

• Invalid API key - the key is wrong or missing from the node's api_key field.
• Images API is not supported for this platform - your key's account group has no image-API access. You need a key from an image-enabled group.
• No available compatible accounts - upstream has no compatible slot right now; retry later.
• Timeouts - with gpt-image-2 keep n=1; the upstream docs recommend not depending on multi-image responses. Banana jobs that overrun report their task id so you can poll it yourself.
• One config trap - success saves nanaix_config.json in the pack folder and prefills future nodes, but a blank visible key still raises an error even when the file holds an old one. And since that file contains your key: never commit it or paste it into a shared workflow. This node phones home with a credential by design - the exact shape of thing the ecosystem has been burned by before - so it's worth the thirty seconds to skim the source before you hand it a key. Small, readable, open codebase, which is more than you get from a lot of API packs.