opensourcefilms-comfyui-workflows

ComfyUI workflow graphs used in production by Open Source Films, released under MIT. These are the actual workflow graphs we use, not vendored copies of other people's tools, plus one custom node package that's genuinely ours.

Unless otherwise stated, original code, workflow files and documentation authored by Open Source Films are licensed under the MIT License. Third-party software, models, weights and datasets remain subject to their respective licences. The MIT licence applied to an OSF workflow or integration does not grant additional rights to its dependencies. Commercial suitability depends on the complete dependency set selected by the user.

What's in here vs. what isn't

• In this repo: the workflow graph JSON files (node connections, parameters, prompts), one OSF-authored custom node package (custom_nodes/mask_propagate_optical_flow/), a pinned installer script, and full dependency documentation.
• Not in this repo: ComfyUI itself, third-party custom node source code, or any model weights. scripts/install-dependencies.sh fetches the third-party custom nodes you need from their own upstream repos, pinned to commits we've verified. Model weights are your own responsibility to download: see the license/source notes in each workflow's .md sidecar, THIRD_PARTY_LICENSES.md, and DEPENDENCIES.lock (machine-readable, canonical if it ever disagrees with the prose docs).

Commercial use

Every workflow is labeled with a commercial_status in its .md sidecar file, one of:

• commercial-compatible; every dependency (ComfyUI core, custom nodes, models) is either permissively licensed or GPL-3.0. GPL-3.0 is commercially usable (it imposes distribution/copyleft obligations, it is not a non-commercial license); we don't treat GPL as a blocker.
• research-only; at least one dependency is CC-BY-NC-4.0 or otherwise restricted to research/education use. Don't use these commercially without separately clearing that dependency with its owner.

Full detail, upstream URLs, and pinned commits/versions for every dependency: THIRD_PARTY_LICENSES.md.

This is our own assessment, based on reading the actual license text of each dependency (not just trusting badges/labels); it is not legal advice. Verify independently before any commercial distribution.

Workflows

workflows/commercial-compatible/

Workflow	What it does
sam3_custompkg_probe.json	Minimal SAM 3 segmentation smoke test
step4_5_sam3_mask.json	SCAIL-2 pipeline: SAM 3 mask generation from a driving video
step6_animation_mode.json	SCAIL-2 pipeline: character animation / motion transfer
step7_replacement_mode.json	SCAIL-2 pipeline: character replacement
gimm_sgm_vfi_video_interpolate.json	Frame interpolation with GIMM-VFI + SGM-VFI
mm_upscale_video_workflow.json	Video upscale with Florence2 auto-captioning
wan22_fmlf_first_middle_last_frame.json	Wan 2.2 first/middle/last-frame to video
video_wan2_2_14B_flf2v.json	Wan 2.2 14B first-last-frame to video
video_wan2_2_14B_fmlf2v.json	Wan 2.2 14B first-middle-last-frame to video (simple variant)
video_wan2_2_14B_i2v.json	Wan 2.2 14B image to video
video_wan_vace_14B_v2v.json	Wan VACE 14B video-to-video with Canny conditioning

workflows/research-only/

Workflow	What it does	Blocked by
videomama_sam_points_text_matte_optflow.json	SAM3 + VideoMaMa mask-guided matting	VideoMaMa (CC-BY-NC-4.0 / Stability AI Community License)
videomama_split_sam_points_text_matte_optflow.json	Same, segmented for longer videos	VideoMaMa
sam_effecteraser_points_text_matte_optflow.json	SAM3 + EffectErase object removal	EffectErase (CC-BY-NC-4.0)
frame_interpolation_test.json	4-way interpolation comparison	BiM-VFI (research/education only)
tween_interpolate_loop.json	Looping interpolation via ComfyUI-Tween	BiM-VFI (research/education only)

Custom nodes

custom_nodes/mask_propagate_optical_flow/ is Open Source Films' own code: two nodes (MaskPropagateOpticalFlow, MaskPropagateFromFlow) that propagate a single mask across a video's frames using OpenCV Farneback optical flow, so you draw or generate one mask instead of one per frame. No external model weights, no third-party dependencies beyond OpenCV/NumPy/ PyTorch (which ComfyUI already requires). Used by the matting/object-removal workflows in workflows/research-only/. Published here for the first time, MIT-licensed; see its own README.

Install

./scripts/install-dependencies.sh /path/to/ComfyUI

This installs the commercial-compatible custom node dependencies, plus mask_propagate_optical_flow, by default. To also install the research-only dependencies (needed for workflows/research-only/*):

INSTALL_RESTRICTED=1 ./scripts/install-dependencies.sh /path/to/ComfyUI

Then in ComfyUI: Workflow → Open and pick a .json from workflows/.

Run on RunPod

We develop and run these workflows on RunPod GPU pods. For the full walkthrough — base image, GPU/VRAM tiers by workflow family, storage, and a verified "fresh pod → clone → install → open workflow" path — see RUNPOD.md. Short version:

git clone https://github.com/OpenSourceFilms/opensourcefilms-comfyui-workflows.git
cd opensourcefilms-comfyui-workflows
./scripts/setup-runpod.sh /path/to/ComfyUI

Requirements

A ComfyUI installation and an NVIDIA GPU. Requirements vary a lot by workflow; the SAM 3 / interpolation workflows are comfortable on a single mid-range GPU, while the Wan 2.2 14B and SCAIL-2 video workflows want 24GB+ of VRAM. We develop and run all of these on GPU pods from Runpod, in collaboration with them on GPU compute: their on-demand pricing makes it practical to spin up a large-VRAM card just for a session rather than owning one outright, which is how we test the heavier workflows here.

License

Workflow graphs, custom_nodes/mask_propagate_optical_flow/, and scripts/install-dependencies.sh: MIT, see LICENSE. Third-party custom nodes and models referenced (not bundled) by these workflows keep their own original licenses; see THIRD_PARTY_LICENSES.md and DEPENDENCIES.lock.

Contributing

See CONTRIBUTING.md.