UniAPI Model Call

One node, one job: UniAPI Model Call is a gateway that lets a ComfyUI workflow call any third-party image service that speaks the OpenAI Images API format - and hands the result back as a real IMAGE tensor you can preview, save, or pipe into more nodes.

Most of your graph is local - checkpoints, samplers, VAE - but every so often the best tool for a job lives in the cloud. Maybe it's a proprietary model you don't have weights for, a hosted model that beats what your GPU can fit, or you just don't want to burn the VRAM on a quick generation. This node is the "bring the API in" bridge: no hand-rolled curl script and copy-pasting images back in, just a node in the graph.

How it works

Under the hood it's two HTTP calls dressed up as a node. In text2img mode it POSTs your prompt to {base_url}/v1/images/generations with ?async=true; in img2img mode it POSTs multipart form data to {base_url}/v1/images/edits, converting up to four IMAGE tensors to PNG along the way. If the service is async and returns a task_id, it polls {base_url}/v1/images/tasks/{task_id} every 10 seconds, up to 60 attempts - ten minutes - with a native ComfyUI progress bar. If the service answers synchronously with a data array, it just decodes that: b64_json gets base64-decoded, url responses get downloaded, and everything gets batched into one tensor.

The resume trick is the genuinely nice touch. Because the poller is synchronous, a long async job can outlast your patience or your internet. The node emits task_id; feed it back into the task_id input and it skips the submit entirely and just queries that task. Disconnected mid-generation? Plug the ID back in and carry on.

The inputs that matter

Only four are required:

• api_key - the bearer token for the service. It lives in the workflow, so treat it like a secret (more below).
• base_url - the API root. Give it https://your-provider.com, not the full /v1/images/generations path; the node appends that itself.
• prompt - your text prompt, multiline.
• mode - text2img or img2img.

The rest are optional. model is the string your provider expects, like gpt-image-1; with many services you can leave it blank and get the default. n generates 1–4 images and batches them into one tensor. seed has a catch: 0 means "random" and is omitted from the request - anything above zero gets sent through. size, quality, background, output_format, and moderation pass through as-is, so they only do something if your provider honors them. response_format defaults to b64_json; flip it to url if your API only hands back URLs (the node still downloads them into a tensor either way).

Outputs

• image - the IMAGE tensor, batched when n > 1. Wire it into a preview or Save Image node.
• image_url - the first image's URL (empty for base64 responses).
• task_id - save this string to resume async jobs later.
• response - the full JSON as text, with status, model, and image count. Your debugging friend.

Install

No model downloads, no dependency hell - the README is explicit that it needs only torch, Pillow, and requests, all already in ComfyUI's base environment. Two routes:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaojuzi2024/ComfyUI-UniAPI.git

then restart ComfyUI. Or search "ComfyUI-UniAPI" in ComfyUI Manager and install from there.

Where people get burned

• It's a network node. Your prompt, your api_key, and in img2img mode your actual images leave your machine for whoever runs the API. ComfyUI embeds workflows into output PNGs by default, so a shared workflow means a shared key - use a scoped key you can revoke.
• A blank white image is the failure signature. The node doesn't throw - it returns a white 1024×1024 placeholder and stuffs the error into the response output. Read that before tearing your hair out. A red placeholder means the task itself failed; a small yellow one means the 10-minute poll timed out (resume with task_id).
• base_url format. Appending the full endpoint path instead of the root breaks every request.
• Not every "OpenAI-compatible" provider is equal. Some skip the async tasks endpoint, some never implemented /v1/images/edits at all. img2img needs an edits endpoint that actually exists - check the provider's docs first.
• Long async jobs hold the queue. The poller is synchronous, so a slow provider ties up a queue slot for up to ten minutes. Fine on a personal box, annoying on shared setups.

The pack is young and quiet - essentially no community footprint yet - so the README and the small, readable source are the docs.