AI/ML API Image Generator (Nano Banana Pro / Seedream 4.5)

Here's the pitch in one line: this node lets you generate images with Google Nano Banana Pro or ByteDance Seedream 4.5 from inside ComfyUI, without downloading a single model file or caring how much VRAM you have. Type a prompt, pay per image, done. That's the whole reason the pack exists - it's a thin wrapper around AI/ML API's hosted image endpoint, and this is the text-to-image entry point of the five-node suite.

It's worth being upfront about what you're getting. Nano Banana Pro is Google's Gemini 3 Pro Image under the hood - the model that does 4K-native output, real text rendering, and "thinks" about the prompt before generating. Seedream 4.5 is ByteDance's flagship, and if you've followed the company's posture, you know it's API-only: it ships nowhere as open weights. So this node is your only realistic way to run these two inside a ComfyUI graph. The flip side is you're renting them per image, and quality/censorship is entirely out of your hands - no negative prompts, no CFG fiddling. Treat it as a paid camera, not a sampler you tune - no KSampler, no scheduler games, no ControlNet in your future.

How it works. Simple under the hood: the node POSTs your prompt to https://api.aimlapi.com/v1/images/generations, then downloads the image URL the API returns and hands you a normal IMAGE tensor. Unlike the pack's video nodes there's no polling - it's one synchronous HTTP call, so the whole generation happens inside a single node execution. The output drops straight into your existing workflow: Save Image, preview, VAE, anything that takes an IMAGE.

The inputs that matter. The trap for beginners is that the parameters are split by model, and the node quietly ignores what doesn't apply. Pick the model first, then look:

• google/nano-banana-pro (default) - uses aspect_ratio, resolution (1K/2K/4K), and num_images (1–4). seed and guidance_scale are ignored.
• bytedance/seedream-4-5 - uses seed (-1 = random), guidance_scale (default 3.5), and watermark. The aspect-ratio/resolution pickers still exist but get baked into a pixel size string server-side.

api_key is the one thing you must fill in - it's a masked password widget, grabbed from https://aimlapi.com/app/keys. The only output is images, so keep your save node attached or the result vanishes into the void.

Installing it. No model files, no heavy deps - the actual requirements are just requests, pillow, and numpy (torch is already in ComfyUI's environment). Easiest path is ComfyUI Manager, searching comfyui-aimlapi-custom-nodes, and letting it clone:

cd ComfyUI/custom_nodes
git clone https://github.com/konradbjk/comfyui-aimlapi-custom-nodes

Restart ComfyUI and you'll find the five nodes under an "AI/ML API" category. The README also ships install_macos.sh / install_windows.bat if you prefer a one-shot installer that copies files and pip-installs the deps into your venv.

What usually bites people. Watch for API request failed with status errors - the node annotates them helpfully: 401 means a bad key, 403 means the key is real but short on credits or expired, 429 is rate limiting. All of those are "check your account" problems, not graph problems. And remember the per-image billing: 4K Nano Banana Pro outputs cost real money (on Google's own API that's roughly 0.24 an image at the top end), so a num_images of 4 is a deliberate budget decision, not a free bonus.