ComfyUI-Wylon

<div align="center"> <h1>ComfyUI-Wylon</h1> <p><strong>Wylon image and video generation, connected to your ComfyUI workflows.</strong></p>

English · 简体中文

Quick start · Example workflows · Usage guide · Report a bug

</div>

Generate images with Qwen and videos with MiniMax H3 through the Wylon API, using native ComfyUI nodes. Connect results to the rest of your graph, keep video soundtracks, and retrieve accepted video tasks after interrupted waiting. No local model weights are required.

• 🖼️ Image generation: Qwen-Image-2512 and Qwen-Image-2.1, with model-specific size presets.
• 🎬 Video generation: text, first/last frames, or image/video/audio references through H3 and Ref2VA.
• 🔌 Native media: IMAGE and VIDEO outputs, native IMAGE/AUDIO reference inputs, plus task IDs and local file paths.
• 🔄 Task recovery: a task panel and two Get Task Result nodes retrieve existing tasks without submitting a new generation.

Generation uses your Wylon account and is subject to its model access, usage charges and limits.

🧭 Choose your integration

| Integration | What it provides | Host / installation | | --- | --- | --- | | 🔌 ComfyUI-Wylon — this repository | Six Python nodes, API key configuration and a task panel | Install in ComfyUI v0.44.0 | | 🎞️ Toonflow-Wylon | A standalone wylon.ts provider | Install in Toonflow v2.0.10; no ComfyUI required | | 🧩 Wylon ComfyTV workflows | Five workflows with paired ComfyTV presets | Optional pack; requires ComfyTV v1.13.1 and ComfyUI-Wylon on the same ComfyUI server |

The ComfyUI runtime ZIP includes the five generation examples and two existing-task retrieval examples below. The optional ComfyTV pack is distributed separately; its installation guide explains where to copy the workflow/preset pairs.

📦 Install

You need a working ComfyUI with the native V3 node/media API, Python 3.10+, a Wylon API key and network access to Wylon and its result hosts. Host version: ComfyUI v0.44.0.

| Platform | Installation method | | --- | --- | | Windows 10/11 x64 | Manual Python/venv or official ComfyUI Windows portable | | macOS | ComfyUI source installation using its Python environment |

Use the Python interpreter that runs ComfyUI. The plugin ships compiled frontend files in web/; ordinary installation needs neither Node.js nor npm. Dependencies are aiohttp, Pillow, numpy and PyAV; ComfyUI supplies PyTorch and its host dependencies.

Manual Python / virtual environment installation

On macOS/Linux, activate your ComfyUI environment and run from the ComfyUI root:

git clone https://github.com/wylon-cloud/comfyui-wylon.git custom_nodes/ComfyUI-Wylon
python -m pip install -r custom_nodes/ComfyUI-Wylon/requirements.txt

On Windows 10/11 x64 with a manual/venv installation, run from the ComfyUI root in PowerShell. This example uses a virtual environment named .venv; substitute the interpreter that starts your host if its path differs.

git clone https://github.com/wylon-cloud/comfyui-wylon.git .\custom_nodes\ComfyUI-Wylon
.\.venv\Scripts\python.exe -m pip install -r .\custom_nodes\ComfyUI-Wylon\requirements.txt

Official Windows portable installation

Run these commands in PowerShell from the portable distribution root, the folder containing python_embeded and ComfyUI:

git clone https://github.com/wylon-cloud/comfyui-wylon.git .\ComfyUI\custom_nodes\ComfyUI-Wylon
.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-Wylon\requirements.txt

ZIP installation

1. Download the source ZIP.
2. Extract it, rename comfyui-wylon-main to ComfyUI-Wylon, and place that folder directly inside ComfyUI/custom_nodes/. The result must be ComfyUI/custom_nodes/ComfyUI-Wylon/__init__.py.
3. Run the dependency command above for your installation, then restart ComfyUI and refresh the browser page.

A locally built ComfyUI-Wylon-0.3.0.zip already contains the correctly named plugin folder. To build that runtime archive and the separate Wylon-ComfyTV-Workflows-0.3.0.zip from a source checkout, run npm run package (developer tooling) or python scripts/package.py with the selected Python; see CONTRIBUTING for build and validation steps.

🔑 Configure your API key

1. Obtain a Wylon API key from Wylon.
2. Open Wylon → Wylon tasks in ComfyUI, or click Wylon configuration / API Key at the bottom of a Wylon generation/result-retrieval node.
3. Enter the key and click Apply API key. The panel shows the effective credential source and a separate connection check.
4. Optionally enable Remember on this machine (private ComfyUI user configuration) to retain the key after restarting.

The key stays on the server and is excluded from workflow inputs and generated metadata. Configuration applies to the ComfyUI server process, shared by users of that server. Opt-in persistence stores plaintext at <ComfyUI user directory>/wylon/private/credentials.json; use it on a trusted host.

For a server-managed setup, set WYLON_API_KEY before starting ComfyUI. For a Windows manual/venv installation, run from the ComfyUI root:

$env:WYLON_API_KEY = 'your-key'
.\.venv\Scripts\python.exe main.py

For official portable, set the same variable from the portable distribution root, then launch your usual script, for example .\run_nvidia_gpu.bat. The variable applies to processes started from that PowerShell session.

Effective priority is session → saved local → environment. The default API base is https://api.wylon.cn/v1. See credential setup, environment commands and removal for details.

🚀 Your first generation

1. Load the Qwen-Image-2.1 example in ComfyUI, or add Wylon Generate Image from the Wylon node category.
2. Select a Qwen model and enter a prompt. Image generation is synchronous; use server_default or a size supported by the selected model.
3. Connect IMAGE to native Preview Image or Save Image, then queue the workflow.
4. For text-to-video, load the H3 example, use its fixed MiniMaxAI/MiniMax-H3 model, and connect VIDEO to native Save Video.

Image generation always uses synchronous POST /v1/images/generations and returns the finished image without a task ID. Both video generation nodes submit an asynchronous Wylon task and keep waiting for the finished media. There is no separate submit-only node.

Downloaded results are stored under <ComfyUI output directory>/wylon/ (normally ComfyUI/output/wylon/). Native Save Image / Save Video can save additional copies. The Wylon tasks panel shows task information and recent local results.

🧪 Example workflows

Open a JSON file in ComfyUI or drag it onto the canvas. Supply your own media for connected loader nodes; example media and API keys are not bundled.

| Workflow | Start here for | | --- | --- | | Qwen-Image-2512 | Text-to-image with Qwen-Image-2512 | | Qwen-Image-2.1 | Text-to-image with Qwen-Image-2.1 | | H3 text-to-video | Video from a text prompt | | H3 first/last frames | Video guided by dedicated frame inputs | | Ref2VA references | Image/video/audio references; provide an image or video and disconnect unused loaders |

For an existing task ID, use Get image task result or Get video task result. The image retrieval example accepts historical asynchronous image task IDs; current image generation cannot create those tasks. Paste the existing ID, then queue; these examples are separate from the five ComfyTV graph/preset pairs. The retrieval node can also run on its own or send native IMAGE/VIDEO into other nodes.

🧠 Models and native ports

| Model | Supported generation in this integration | Inputs / limits | | --- | --- | --- | | Qwen/Qwen-Image-2512 | Text-to-image | Prompt and model-specific size; one image per request | | Qwen/Qwen-Image-2.1 | Text-to-image | Prompt and model-specific size; one image per request | | MiniMaxAI/MiniMax-H3 | Text-to-video, first/last-frame video | Optional native first_frame / last_frame, one image each | | MiniMaxAI/MiniMax-H3-Ref2VA | Video with image/video/audio references | Up to 9 images, 3 video URLs, 3 audio clips; requires at least one image or video |

Video output is 768p, 5–15 whole seconds. Ref2VA does not accept the dedicated first/last-frame roles; audio alone is insufficient. Reference videos must be public HTTPS URLs reachable by Wylon, entered one per line in the urls control of Wylon Reference Video URL and connected to Ref2VA’s reference_videos port. Local paths, Base64 video, localhost/LAN URLs and automatic uploading are unsupported.

| Node | Stable node ID | Native media ports | | --- | --- | --- | | Wylon Generate Image | WylonGenerateImage | IMAGE output | | Wylon H3 Video | WylonH3Video | IMAGE frame inputs → VIDEO output; fixed H3 | | Wylon Ref2VA Video | WylonRef2VAVideo | Native IMAGE/AUDIO references + dedicated reference-video bundle → VIDEO output; fixed Ref2VA | | Wylon Reference Video URL | WylonReferenceVideoURL | Public HTTPS URLs → WYLON_REFERENCE_VIDEO reference bundle | | Wylon Get Image Task Result | WylonResumeImageTask | Existing task_id → IMAGE output | | Wylon Get Video Task Result | WylonResumeVideoTask | Existing task_id → VIDEO output |

Wylon Reference Video URL only validates and packages URLs; it does not download, upload or generate media, read an API key, or need a configuration button. Empty urls produces an empty reference bundle, so the same graph can generate from images only. Ref2VA still requires an image or video reference. Its dedicated WYLON_REFERENCE_VIDEO connection is separate from native VIDEO, which carries a generated local video result; connecting a local video output does not upload it. Ref2VA has no URL text control.

Each video generation node has one fixed model and exposes only its supported inputs: H3 for text/frames, Ref2VA for references. Neither has a model selector. Generation/result-retrieval outputs include task_id and file_path; image generation keeps task_id empty for compatibility and has no task-based recovery. Native VIDEO retains the downloaded MP4's soundtrack.

For video requests, a nonnegative seed is forwarded within the supported API range; -1 omits it. Model support determines its effect, so a fixed value is not a promise of identical output. control_after_generate changes the local video seed for the next queue run; it is not an API parameter. Image generation has no mode, seed or seed-policy controls.

See models and media for image sizes, clip duration/size limits and native video component connections. Wylon's service documentation covers image generation and video generation.

🔄 Recovery and troubleshooting

Generation POSTs are never retried automatically. For video or an existing historical image task, after a timeout, interruption or download failure, copy the accepted task ID from the panel/error into Wylon Get Image Task Result or Wylon Get Video Task Result. Get Task Result uses GET to query the same task and download its result; it does not create, restart or cancel cloud tasks.

A failed generation node cannot pass its task_id output downstream, so copy the ID manually from the task panel or error into the retrieval node.

Stopping local waiting does not cancel cloud work. If creation times out without a task ID, check Wylon history and usage before resubmitting. Requeueing a Generate node sends a new generation request and may incur another charge.

| Symptom | First check | | --- | --- | | Nodes or Wylon menu missing | Correct folder nesting, host version and dependencies in ComfyUI's Python; restart and refresh | | Key rejected or unexpected account | Effective key source in the panel; session/saved keys override the environment | | Reference validation error | Fixed node model, connected roles, media counts and public HTTPS video URLs | | Download or codec failure | Check network, disk space and PyAV; if an existing task ID is available, preserve it and use a Get Task Result node while its result remains available |

More: recovery and troubleshooting · update and uninstall.

🤝 Contribute and get help

Use GitHub Issues for bugs and installation problems. Include host/plugin versions, the workflow name, reproducible steps and a redacted error. For sensitive reports, follow SECURITY.md.

Contributions are welcome: development and checks · contributor checks · changelog.

⭐ Community

<div align="center">

Page views · powered by MoeCounter

</div>

📄 License

MIT · Third-party notices