Replicate Custom (beliebige Modell-ID)
Replicate Custom — the escape hatch for the other 400,000 models
- image
- raw_output
- image
Replicate Select ships a curated catalog of two dozen nodes, but Replicate hosts a lot more than two dozen models. That gap is what ReplicateCustom is for: the author calls it the escape hatch, and that's exactly right. You type any Replicate model ID - owner/model - paste its inputs as JSON, and the node calls it like one of the curated ones. If the pack hasn't wrapped the model you want, this is the door.
It's the second most useful node in the pack after whatever you actually came for, and it's the reason the pack can stay "schlank statt Katalog-Flut" (lean instead of a catalog flood, per the README) - the author keeps the curated list small on purpose because Custom catches everything else.
How it works
Two required inputs and one optional, then it's pure plumbing:
- model_id - the Replicate model string, e.g.
bytedance/seedream-4.5. Owner/model format. - input_json - the model's input parameters as JSON. The default is
{ "prompt": "" }, but every model expects different keys, so you'll edit this per model. This is where the schema knowledge lives: you need to know what the model you're calling actually takes. - image (optional) - an IMAGE tensor. If set, it's appended to the input as the
imageparameter (handy for img2img-style models).
The node parses your JSON, calls replicate.run(model_id, input=args), and returns two outputs: raw_output (STRING - the model's raw response) and image (an IMAGE tensor, populated if the output was a PNG/JPEG). If the output isn't an image, you get raw_output plus a byte-count note, and the image output is a dummy tensor.
The honest tradeoffs
This node trades the curated nodes' safety for reach. The curated catalog nodes have typed inputs, tooltips, and sane defaults - the author already did the schema work. ReplicateCustom hands that responsibility to you: mistype a key name, and the call fails or silently misbehaves. You'll want the model's Replicate page open in another tab to copy its exact input names. There's also the security frame that applies to every API node in this ecosystem: it sends your data and your key's calls to a third party by design, so only point it at models you actually recognize.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select
Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Auth via environment variable:
export REPLICATE_API_TOKEN=r8_...
Where people get burned
- JSON must be valid and match the model's schema. The most common failure is a model that doesn't have a top-level
promptfield, or an enum value you mistyped. Check the model's schema page before pasting. - Output type is a guess.
raw_outputis always a string;imageonly fills in for image outputs. Don't assume every model's output becomes a usable IMAGE. - It's a per-call meter. Just like the curated nodes, every run costs money - and with no tooltips guiding you, it's easy to fire off a dozen wrong-shaped calls while you're learning the model's input. Read the schema first.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model_id | STRING | owner/model | z.B. bytedance/seedream-4.5 |
| input_json | STRING | { "prompt": "" } | Input-Parameter als JSON |
| imageopt | IMAGE | Wird als 'image'-Parameter angehängt, falls gesetzt |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| raw_output | STRING | — |
| image | IMAGE | — |