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

Local ComfyUI editing is a ritual: load the image, paint a mask, pick an inpaint model, fight the denoise slider, hope the edges blend. This node skips the entire ritual. You feed it an image and a sentence - "turn this into autumn with falling leaves" - and Nano Banana Pro Edit or Seedream 4.5 just does it. No mask, no latent round-trip, no VRAM budget. If you've ever wanted the "describe the change" editing that hosted tools have, without leaving your ComfyUI graph, this is that.

It's the image-to-image sibling of the pack's AIMLAPIImageGenerator, and mechanically it's nearly identical: same https://api.aimlapi.com/v1/images/generations endpoint, same synchronous call-and-download. The difference is the payload. Your IMAGE tensor is serialized into base64 PNG data URIs and sent up as image_urls - up to 14 of them, so batches are genuinely supported - and the model edits from those instead of from nothing. The node then downloads the result and hands you back a fresh IMAGE tensor.

Inputs that matter. The same model-split gotcha as the generator applies, so choose your model first:

• google/nano-banana-pro-edit (default) - respects aspect_ratio, resolution (1K/2K/4K), and num_images (1–4).
• bytedance/seedream-4-5 - respects seed, guidance_scale, and watermark; the ratio/resolution pickers become a size string.

The three you'll actually touch day to day are api_key (masked password widget, from https://aimlapi.com/app/keys), prompt, and image. Don't over-think the prompt - the default text ("Transform the scene to autumn...") is just the author's demo line, not guidance; describe the change you want in plain language. Note this is prompt-driven editing, not strict instruction-following: it reinterprets the whole image, so it's great for relighting, restyling, or swapping a subject, and less reliable for "change exactly this one region" surgery.

The single output, images, plugs straight into any Save Image or preview node, and because it's a normal IMAGE tensor you can chain these - generate in this pack, edit the result, then keep going downstream. That's the genuinely nice part: the cloud models feed your local pipeline instead of living in a separate tab.

Installing it. Same story as the rest of the pack - ComfyUI Manager, search comfyui-aimlapi-custom-nodes, or:

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

Restart and look for the "AI/ML API" category. Dependencies are just requests, pillow, and numpy (torch ships with ComfyUI), and there are zero model downloads - the work happens on AI/ML API's servers, per-image, for money.

What usually bites people. The error text is your friend: API request failed with status 401/403 points at the key or its credits, 429 at rate limits. Also worth knowing: because your image travels as base64, a big batch of 4K inputs gets heavy - the encoding adds ~33% overhead on top of the PNG, so runs get slower as your input gets larger. That's not a bug, it's the physics of sending images to a remote API. And since both models are closed and API-hosted, don't expect the reproducibility of local seeding - this is a tool for "I need this one image changed," not for parameter sweeps.