🤖 Replicate 萬用生成 / Universal Generator (All Models)

The 🤖 Replicate 萬用生成 / Universal Generator is the node that finally lets you skip the search for a local Nano Banana checkpoint. Stop looking - there isn't one. Sora 2, Veo 3.1, Nano Banana Pro, Kling, Seedance: none of them have open weights, so "run it locally" is off the table at any VRAM budget. This node is the other door. You pick a model from a dropdown of 34, type a prompt, and the node packages everything up and sends it to Replicate's cloud, then drops the result back into your graph as native VIDEO, IMAGE, or AUDIO - as if a local sampler had made it.

This is the API-wrapper pattern in its cleanest form: a ComfyUI node that is really an HTTP client with your key baked in. The official ComfyUI Partner Nodes do the same job on prepaid credits through your Comfy account; this one instead uses your own Replicate token and Replicate's pay-per-call billing. Same closed models, different meter, and it also runs open models (FLUX schnell, Wan, Stable Video Diffusion) on someone else's GPU if yours can't cope.

How it works

The model dropdown is the only required input - everything else is optional and, crucially, inert when the selected model doesn't use it. ComfyUI can't hide parameters on a dropdown node, so all ~30 of them sit there, but the node filters the ones each model needs before the API call. You don't need to know whether Veo wants a "guidance_scale" or a "duration"; it just picks the relevant ones.

The one input you'll actually touch beyond prompt is images - a unified socket that auto-routes to whatever the model calls its image parameter: Veo/Wan get image, Sora gets input_reference, MiniMax gets a first frame, Kling gets a start image, and Nano Banana/Seedance get a full reference-image list. Wire the pack's Multi Image Input node in here and Nano Banana can see all 14 references in one call. last_frame stays separate, for Veo's end-frame editing. A video + audio pair handles the lipsync models.

Under the hood there's real engineering, not just a token request. Predictions are created with wait=False and then polled, so a five-minute Veo render won't trip a read timeout. Results download to ComfyUI's temp directory, and the outputs come out as native types: video → Save Video, image → Save Image, audio → Save Audio. Backend ffmpeg muxes models that return separate video and audio streams back into one file, and a 1-second debounce means double-clicking Run doesn't bill you twice. file_path carries a .glb path for the 3D model, info carries execution details.

What to watch

Cost. This is metered per call, and the video jobs are where a session gets expensive fast - a handful of 4K Veo clips can outspend a month of Comfy Cloud. It's the real price of models that refuse to run locally, not a bug. Also: your prompts and reference images leave the machine, and the closed models enforce their own content filters - there are no weights to abliterate.

Install

ComfyUI Manager, search Replicate API NM, or:

cd ComfyUI/custom_nodes
git clone https://github.com/neurmostudio0409/ComfyUI-replicate-api-NM
cd ComfyUI-replicate-api-NM
pip install -r requirements.txt

Then create a .env in the pack folder with your token from https://replicate.com/account/api-tokens (the pack reads it on load), and make sure FFmpeg is on your PATH - it's the system dependency for audio muxing (choco install ffmpeg, brew install ffmpeg, or apt install ffmpeg). Python deps are just replicate, requests, opencv-python, soundfile, scipy - nothing exotic, and torch is ComfyUI's.

If you get a token error, check the .env format (export REPLICATE_API_TOKEN=r8_...); a "model parameter" error means a required input is missing and the node's preflight now catches it before it bills you. Old Veo timeouts and Grok image-format errors were fixed in v2.5.x - if you hit those, update. It's a small pack (this page has essentially zero search impressions yet), but it's GPL open source, on the Comfy Registry, and ships 21 unit tests with CI. Any node that holds your API key and phones home by design is worth eyeballing before first run - that's the category's one honest caution, not a strike against this one.