Skip to content

Beakfx/BCE

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BCE — Batch Comfy Extensions

Bring ComfyUI workflows into Autodesk Flame's Batch.

BCE nodes in a Flame Batch schematic — Outpaint, Inpaint, and SAM, with the Inpaint note open

BCE is a whole system for using and making Batch nodes that do Comfy workflows. A BCE node reads parameters in our familiar UI, patches a ComfyUI workflow template with those values, and renders through your chosen backend — local Comfy, a LAN render node, or Comfy Cloud. The result imports back into Flame as a cached clip. All your Comfy settings, including seeds, are saved in your Batch setups.

If you've ever wanted to drop a ComfyUI workflow into the middle of a Batch tree without leaving Flame, this is that.


What's in the box

Three node categories, each with several templates, covering a class of Comfy workflow:

  • BCE Outpaint — extend single frames
  • BCE Inpaint — replace or remove content using a matte
  • BCE SAM — use Segment Anything Models to generate mattes. Better than the Semantic keyer and faster than whatever Magic Mask thing you were doing.

Three render backends, switched with a toggle in Setup:

  • Local — ComfyUI on your workstation
  • LAN — ComfyUI on a render node, via a shared filesystem
  • Cloud — Comfy Cloud (account + API key)

You can use more than one — local for fast iteration, LAN if you have a render node, or cloud for offsite rendering. Cloud is good for when your GPU is busy, and great if you're on a Mac.


Why BCE

A few things BCE does differently than the typical "drop a PNG into Comfy" workflow:

  • Imports stay. BCE forces MediaHub cache mode during import, so every result lands as a fully cached clip in Flame. Job folders on disk are safe to delete — your timeline keeps the media. No "missing source media" surprises after a purge.
  • Deterministic and archive-safe. Every parameter — prompt, seed, sliders, every Matchbox value — lives in the Flame Batch setup. The prompt itself is stored in the BCE node's Note (to finally put that to good use). Pull a Batch setup out of archive six months later, re-render, get the exact same result.
  • 16-bit float EXR end-to-end on local and LAN. Source plates go to Comfy as EXR and results come back as EXR — no PNG round-trip, no 8-bit quantization between Flame and inference.
  • Cloud uses 16-bit RGBA TIFF transport. Comfy Cloud doesn't currently support EXR LoadImage, so BCE uploads sources as 16-bit TIFF and gets EXR back. It's a kludge, and it's named as one — when cloud EXR support lands, BCE will switch.
  • Color space is preserved. The clip BCE imports back into Flame wears the color space tag of the source clip — tested across project defaults plus oddballs like BMD Film and RED Dragon. Comfy's side of the pipeline is still yours to manage in your templates.
  • Three backends, one workflow. Switch between local, LAN, and Comfy Cloud with a toggle in Setup. The same BCE template runs on all three (with some node compatibility caveats — see workflows.md).
  • No black-box patching. Every shipped template is a normal ComfyUI JSON. The [BCE:*] tag contract is the only thing BCE cares about; everything else is yours to inspect, change, or replace.

The trade-off: BCE constrains a workflow to what its Matchbox UI exposes. For one-off experiments and workflows BCE doesn't cover, fall back to ComfyUI directly. BCE is for the workflows you want to drive from Flame.


Install

Download the release zip, unzip, run the installer:

./install_bce.sh

Then BCE → Setup and Config in Flame.

See install.md for the full walkthrough, including picking and configuring a backend.


Documentation


Requirements

  • Autodesk Flame 2023.1 or later, Linux or Mac
  • One backend: a local NVIDIA GPU, a LAN render node, or a Comfy Cloud account
  • Video nodes (SAM + future video nodes): ffmpeg on the Flame workstation — see install.md for install commands

Approach

BCE treats GenAI as plate generation, not procedural reproducibility. You're rendering an image, not building a deterministic pipeline node. Iterate, pick a take, move on.


Open by design

BCE is built to be extended, not just used. Three places the door is deliberately left open:

Your own workflows. Every shipped template is a normal ComfyUI JSON. Open it in Comfy, modify it, retag the nodes with [BCE:*], drop it back into BCE. The tag contract is the only thing BCE cares about — everything else is yours to change.

Your own templates. Pick a different model, swap a sampler, change resolution behavior, add a control net. Save the JSON to ~/bce/templates/ and point your BCE node at it. No code changes needed.

Your own node categories. When a workflow is shaped enough differently that it needs its own UI (video instead of stills, two inputs instead of one), copy an existing bce_<node>.py, edit the constants at the top, build a matching Matchbox UI, and you have a new BCE node type. The four-file pattern (bce_<node>.py, bce_launch.py, bce_lib.py, bce_runner.py) keeps "the file you edit" small. See CLAUDE.md and GLOSSARY.md for the contract details.


Credits

  • ComfyUI, the engine BCE drives.
  • The Logik community, especially Michael V (pyflame.com) for the Flame Python style and reference scripts that helped define BCE's shape.
  • The Radiance Digital Cinema custom nodes for the EXR pipeline.

Author

Built by Beak for the Flame community. Find me on Logik.


License

See license.

About

BCE - Batch Comfy Extensions for Flame

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages