ComfyUI Extension: ComfyUI Deadline Submission
Run ComfyUI workflows without the setup
No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.
ComfyUI plugin for submitting workflows to Thinkbox Deadline for distributed rendering. Enables render farm distribution with configurable job settings, batch processing, and automatic seed variation.
Looking for a different extension?
Custom Nodes (4)
README
ComfyUI Deadline Plugin
Submit ComfyUI jobs to Thinkbox Deadline from inside ComfyUI.
Quick demo:
<p align="center"> <a href="https://youtu.be/NFmIvEoEPiU"> <img src="https://img.youtube.com/vi/NFmIvEoEPiU/maxresdefault.jpg" alt="ComfyUI x Deadline demo" /> </a> </p>This plugin adds two nodes:
Submit to Deadlinesends the current workflow to the farm.DeadlineSeedgives each Deadline variation a predictable seed.
The submitter does not render the workflow locally. It only packages the job and sends it to Deadline.
Install
Clone this into ComfyUI/custom_nodes:
git clone https://github.com/doubletwisted/ComfyUI-Deadline-Plugin.git
Restart ComfyUI.
Then deploy the Deadline plugin:
powershell.exe -ExecutionPolicy Bypass -File .\scripts\deploy_deadline_plugin.ps1
The deploy script asks Deadline where the repository lives by running deadlinecommand -GetRepositoryPath, then copies plugins/ComfyUI into custom/plugins/ComfyUI.
You can also point it at the repo yourself:
powershell.exe -ExecutionPolicy Bypass -File .\scripts\deploy_deadline_plugin.ps1 -RepositoryPath "\\YOUR-SERVER\Repository\custom\plugins"
In Deadline Monitor, set the ComfyUI plugin's ComfyUI Installation Paths. Put one portable ComfyUI root per line:
C:\ComfyUI_windows_portable
D:\Apps\ComfyUI
\\YOUR-SERVER\software\ComfyUI
Workers try those paths in order and use the first one that contains ComfyUI\main.py and python_embeded\python.exe.
Use It
Add Submit to Deadline to your workflow, set output_directory to a path the farm can see, and run the workflow in ComfyUI.
Use batch_count for how many variations you want. Use chunk_size for how many variations a Deadline task should process before it finishes. These are variations, not animation frames.
If you want seeds to change per variation, use DeadlineSeed. A base seed of 1000 becomes:
variation 0 -> 1000
variation 1 -> 1001
variation 2 -> 1002
Input Files
Normal ComfyUI loader nodes copy pasted or uploaded files into ComfyUI/input. That folder usually exists only on the machine where you submitted the job, so the plugin stages those referenced files next to your output folder:
<output parent>\input\
Workers start ComfyUI with that folder as --input-directory.
Only files used by the submitted prompt are copied. If a file already exists and is identical, it is reused. If the name collides with a different file, the plugin gives the staged copy a unique suffix.
Absolute path loader nodes are left alone. Those paths must already be valid on the farm.
Workflow Metadata
Deadline gets both files:
prompt_to_execute.jsonis the API prompt the worker renders.workflow.jsonis the normal ComfyUI workflow with node positions.
The worker writes the normal workflow metadata back into the output image when ComfyUI provides it. So dragging the finished image back into ComfyUI should reopen the readable graph, not the ugly API-format graph.
Maintenance Scripts
There are sanitized Deadline maintenance-job templates in scripts/maintenance.
Copy them somewhere your workers can reach, edit the default paths, then submit sync jobs with:
python \\YOUR-SERVER\share\scripts\maintenance\submit_comfy_sync.py --type both
You can also run only one side:
python \\YOUR-SERVER\share\scripts\maintenance\submit_comfy_sync.py --type installation
python \\YOUR-SERVER\share\scripts\maintenance\submit_comfy_sync.py --type models
Notes
- This targets portable Windows ComfyUI workers.
- Deadline handles render timeouts. Set those in Deadline Monitor.
- The old fake
/deadline/*API routes are gone from this plugin. If you useComfyUI-Deadline-Distributed, it owns its own routes.
Run ComfyUI workflows without the setup
No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.