Liquid Inference

Liquid Inference for ComfyUI

This package adds one node to ComfyUI: Liquid Image (paid API). The node sends a text prompt to Liquid Inference, an auction exchange for inference. Sellers post offers for image models. The exchange gives your request to the best offer that is under your price cap, and you pay for the images that you get.

[!IMPORTANT] Each run of the node is a paid request to an external service. Your account pays for it. The prompt goes to the seller that wins the auction.

Requirements

• ComfyUI.
• A Liquid Inference API key with the spend scope, and a balance on the account.

Install

1. In ComfyUI, open Manager, then Custom Nodes Manager.
2. Search for "Liquid Inference".
3. Select Install, then restart ComfyUI.

To install by hand, clone this repository into ComfyUI/custom_nodes/, then restart ComfyUI. The package has no Python dependencies to install.

Set the API key

1. Open Settings, then Liquid Inference.
2. Paste your key into API key.
3. Select Save.

The key goes to the ComfyUI server. The server keeps it in a file in the ComfyUI user directory, user/__liquid_inference/api_key. On Linux and macOS, only your system user can read this file. ComfyUI does not serve folders whose names start with __.

As an alternative, set the LIQUID_API_KEY environment variable before you start ComfyUI. The environment variable has priority over the stored key.

The node has no input for the key. ComfyUI writes each input value into the workflow file and into the metadata of each saved image. A key in an input would thus go to each person who gets one of your images or workflows.

Use the node

1. Double-click the canvas and search for "Liquid Image".
2. Write a prompt, and select a model.
3. Connect images to Save Image or to a different image node.
4. Queue the workflow.

The example_workflows folder has three workflows: text to image, text to image with an upscale, and two models side by side. In ComfyUI, open them from Templates.

Inputs

| Input | Description | | --- | --- | | prompt | The text that describes the image. | | model | An image model that a seller offers now. The node gets the list from the exchange and refreshes it each minute. | | size | auto, or a size in pixels. auto lets each seller use its default size, so more offers can take the request. An explicit size goes only to the sellers that declared it. | | n | The number of images, from 1 to 4. | | cap_usd | The most that one run of the node can cost, in USD. |

The cap applies to each request, not to the workflow. A workflow with three Liquid nodes and a cap of 0.05 can cost up to 0.15 USD for each run. If you queue the workflow ten times, each run is a new request with its own cap.

ComfyUI does not run a node again when its inputs did not change. Thus a second queue of the same workflow does not buy the same images again. Change an input to buy new images.

Outputs

| Output | Description | | --- | --- | | images | The images, as a ComfyUI IMAGE batch. | | job_id | The exchange's ID for the job. Give this ID to support with each question about a charge. | | receipt | JSON with the job ID, the model, the seller, the charge in USD, your cap and the node version. |

Privacy

• The prompt goes to the Liquid Inference router, and then to the seller that wins the auction. The seller can be a different company for each run.
• The images come back from the seller. If the seller returns a link and not the image bytes, the node gets the image from that link. The node does not send your key to that link.
• Anyone who can open your ComfyUI server can queue paid runs with your key. Do not start ComfyUI with --listen on a network that you do not trust.

Retries

The node never sends a request again by itself. A second request could buy the images a second time.

If the connection stops before the answer comes back, the node shows an error that says that the outcome is unknown. The job possibly ran, and your account possibly paid for it. The error gives an Idempotency-Key. Give it to support to find the job.

Troubleshoot

| Error | What to do | | --- | --- | | No Liquid Inference API key | Set the key in Settings, or set LIQUID_API_KEY, then try again. | | HTTP 401 | The exchange does not know the key. Make a new key and save it. | | HTTP 402 insufficient_balance | Add funds to your account. | | HTTP 400 unsupported_parameter | No offer for the model accepts a value that you sent, usually size. Set size to auto, or select a different model. | | HTTP 503 no_eligible_offer with cap_exceeds_buyer_cap | Each offer costs more than cap_usd. Increase cap_usd. | | HTTP 503 no_eligible_offer, other reasons | No seller can take the request now. Try again later, or select a different model. | | HTTP 404 model_not_found | No seller lists the model now. Reload the page to get the current model list, then select a model. | | HTTP 503 model_unavailable | The model has sellers, but none of them can take a job now. Try again in a minute, or reload the page and select a different model. | | HTTP 503 provider_failed | The seller failed before it made an image, and you were not charged. Run the node again. If no other seller has the model, the next run gets model_unavailable. | | HTTP 502 or 504 with a web page | A proxy between ComfyUI and the exchange answered. Look at your receipts before you run the node again. | | The model list shows "no image models" | ComfyUI could not get the model list. Check the network and the base URL, then reload the page. | | The outcome is unknown | Do not run the node again at once. Look at your receipts, or give the Idempotency-Key to support. |

Each refusal from the exchange includes the code and the reasons that the offers were excluded. A refusal before the auction costs nothing.

Base URL

The node sends requests to https://router.inference.ai.exchange/v1. To use a different router, set Base URL in Settings or set the LIQUID_BASE_URL environment variable.

Development

The project uses uv.

uv run pytest            # unit tests, with a local stub router
uv run ruff check
uv run ruff format --check
uv run mypy

The smoke test installs a pinned ComfyUI release and CPU torch, starts ComfyUI against the stub router, and runs each example workflow. Then it makes sure that no saved image, no workflow and no server answer holds the key. The first run takes some minutes.

uv run pytest -m smoke

To keep the ComfyUI installation for the next run, set LIQUID_SMOKE_DIR to a folder.

Contribute

This repository is a copy of the source, which is not public. Each change comes from the source, and a tag v<version> publishes that version to the Comfy Registry. Open an issue or a pull request here. A maintainer applies an accepted change in the source, and the next copy includes it. Do not expect a merge of the pull request itself.

License

MIT. See LICENSE.