Removes green and purple fringe from videos, touching just the chroma channels.
Fringe is a shadow cast by a source. First, find all the sources ("casters"). Then, find all the nearby fringe ("shadows"). Then, pull that chroma towards the clean, local tone.
Green Fringe is cast by saturated warm sources (red / purple). Purple Fringe is cast by bright blown highlight sources.
Green fringe can be cast by purple fringe. If you remove purple fringe first, the green fringe it casts will be orphaned (without a caster), and will be unable to be removed by the green fringe algorithm. So, run green first, then purple.
defringe_numpy.py is the canonical numpy implementation; defringe_torch.py is its ONNX-exportable twin (convolutions in place of scipy's label/EDT). They share geometry.py (the resolution-relative → pixel conversion and reach/area calibration) so the two can't drift; tests/ pins the twin to the reference.
Five tabs: load a clip, tune each pass on a colour wheel, check it holds over time, export ONNX.
0 · Source — point at a video, extract a few seconds into memory.
Overlays — see what each pass touches: the casters and their reach, and the correction alpha.
 |
 |
| casters & reach | alpha |
1–2 · Green & Purple — drag the Casters and Shadows wedges; the sliders below mirror the handles.
 |
 |
| green pass | purple pass |
3 · Temporal — scrub corrected frames; the flicker heatmap flags shimmer.
4 · Run — bake settings into a portable ONNX model, run it over the clip or the whole video.
app.py Gradio tuner — layout + event wiring (controllers); the launcher
defringe/ the library package
defringe_numpy.py the domain logic (source of truth): green & purple casts,
the shared cast engine, Lab/soft-step helpers — pure numpy/skimage
defringe_torch.py torch/ONNX twin — tracks the numpy reference
geometry.py resolution-relative → pixel conversion + reach/area calibration,
shared by both twins so they can't drift
parameters.py the tunable-parameter spec: one Param row per knob (name, default,
range, label, help) — the single source the rest projects from
sliders.py builds a Gradio slider per spec; the registry, persistence, profiles
views.py detection overlays + colour-wheel config (presentation)
video_io.py ffmpeg wrappers: decode clips/frames, stream-decode, encode
onnx_runtime.py ONNX Runtime device selection + session building
assets/ frontend: defringe_wheel.js, gradio_ui.js, acc.css
model/ cast_defringe.onnx — exported model (uint8 RGB in/out, dynamic N/H/W)
colab_defringe.ipynb GPU runner: ONNX over a whole video, colour-correct encode
tests/ numpy ↔ torch/ONNX conformance (mean/p99 tolerance)
samples/ sample stills + clip used by the app
docs/ README screenshots; crops/ holds the before/after squares
Needs uv and ffmpeg (brew install ffmpeg, or apt install ffmpeg).
uv fetches a compatible Python (≥ 3.11) and the deps for you — nothing else to install.
# get uv if you don't have it: curl -LsSf https://astral.sh/uv/install.sh | sh
uv run python app.py # tuner at http://127.0.0.1:7862Running the conformance suite (devs only) adds one extra: uv sync --extra test, then pytest.
Use the app to export a tuned algorithm as ONNX, and run it against your video frames per your preference. Optionally, use the colab_defringe notebook to run it against a video.