Guides/Agent integrations/ComfyUI

Using wylon image models in ComfyUI

wylon provides an OpenAI Images-compatible synchronous endpoint. ComfyUI can call it through a third-party node that supports a custom api_base; this guide uses comfyui-gpt-image.

Prerequisites

A working ComfyUI installation.
Permission to install custom nodes and edit node files.
A wylon API key and an image model ID, such as Qwen/Qwen-Image-2512.

1. Install comfyui-gpt-image

In ComfyUI, open Manage Extensions, switch the filter to node packs, search for comfyui-gpt-image, install it, and restart ComfyUI.

warning

The third-party node's size options must be updatedThe node was designed for OpenAI gpt-image-1, whose default sizes differ from the current wylon example model. Follow the next steps to edit the node source and restart ComfyUI.

info

This is a third-party nodeIt is not maintained by wylon. Back up local workflows and modified node files before upgrading ComfyUI or the node.

2. Confirm model parameters

GET /v1/models confirms model IDs and availability; use the API documentation for parameter limits. Qwen/Qwen-Image-2512 currently supports:

Parameter	Values	Notes
`size`	`1328x1328`, `1664x928`, `928x1664`, `1104x1472`, `1584x1056`	Default: `1664x928`
`n`	`1`	One image per request
`output_format`	`png`	PNG output
`negative_prompt`	Supported	Content to exclude

See the synchronous image API for current details. Do not add inference_steps; it is not a supported request field.

3. Add the size options

Edit ComfyUI/custom_nodes/comfyui-gpt-image/nodes_api.py. Find the fixed size options list and add the wylon model sizes without removing the existing values:

"options": [
    "auto", "1024x1024", "1024x1536", "1536x1024",
    "1328x1328", "1664x928", "928x1664", "1104x1472", "1584x1056"
],

Save the file and restart ComfyUI.

update

Node updates may overwrite this changeAfter updating comfyui-gpt-image, check the size options in nodes_api.py again.

4. Configure and run the node

Add a Ports OpenAI GPT Image 1 node and configure it:

Field	Value
`prompt`	A non-empty image description
`api_base`	`https://api.wylon.cn/v1/` (keep the trailing slash)
`auth_token`	Your wylon API key
`model`	A complete model ID, such as `Qwen/Qwen-Image-2512`
`size`	A size supported by the model
`n`	`1` for this example
`quality`, `background`, `moderation`	Keep the default values; actual support depends on the selected model

link

api_base must end with a slashThe node uses urljoin to construct request URLs. Without the trailing slash, /v1 may be replaced and the request will return 404.

Connect the image output to Save Image or Preview Image, then run the workflow.

key

Remove the key before sharing a workflowauth_token may be saved with the workflow. Remove the API key before exporting, uploading, or sharing it, and confirm that the file contains no plaintext credential.

5. Improve download resilience

The node makes a second request to download the signed result URL. If the network is unreliable, edit validate_and_cast_response() in nodes_api.py to add a timeout, integrity check, and bounded retries. First add import time near the top of the file, then replace the single-download logic with:

last_err = None
for attempt in range(4):
    try:
        with requests.get(image_url, stream=True, timeout=(10, 60)) as response:
            if response.status_code != 200:
                raise Exception("HTTP %s" % response.status_code)

            buffer = io.BytesIO()
            for chunk in response.iter_content(65536):
                if chunk:
                    buffer.write(chunk)

            expected = response.headers.get("Content-Length")
            if expected and buffer.tell() < int(expected):
                raise Exception("Truncated: %d/%s" % (buffer.tell(), expected))

        buffer.seek(0)
        img = Image.open(buffer)
        img.load()
        break
    except Exception as error:
        last_err = error
        time.sleep(2 * (attempt + 1))
else:
    raise Exception("Failed to download the image: %s" % last_err)

This change retries only the result download. Once generation has been accepted, do not rerun the whole workflow merely because the download failed; doing so may create another billable request.