vidgod/README.md

VIDGOD

Open-source AI video production toolkit for music-video editing around DaVinci Resolve. All inference is local on Apple Silicon (PyTorch MPS + MLX) — no cloud, no subscriptions.

Built for the workflow: hoard old footage → normalize it → index every shot and every spoken word → search it → arrange to a beat grid in Resolve → cut actors out with clicks instead of rotoscoping splines.

The tools

Command	What it does
bin/vg-roto CLIP --point x,y	Click an actor → tracked cutout as ProRes 4444 with alpha (SAM 2.1 + MatAnyone). --grab-frame saves the prompt frame so you can find coords. --mode mask for hard binary masks via SAM2 video propagation.
bin/vg-transcode DIR	Batch-convert old AVI/WMV/MPEG/etc → ProRes LT, mirrored dir tree, auto-deinterlace, skips already-done files. --h264 for small proxies.
bin/vg-index PATHS	Shot-detect (PySceneDetect) + transcribe (mlx-whisper large-v3-turbo) + thumbnail everything into library/clips.sqlite (FTS5). Resumable.
bin/vg-find WORDS	Full-text search the dialogue; shows the enclosing shot. --cut DIR exports each matching shot as a ProRes clip. --shots VIDEO lists a video's shots.
bin/vg-beats SONG	Beat-track a music file (librosa) → .beats.csv + a Resolve marker EDL (import: right-click timeline in media pool → Timelines → Import → Timeline Markers from EDL).

Install

Needs: Apple Silicon Mac, brew install ffmpeg uv, ~4GB disk for venvs + models.

git clone ssh://git@100.71.119.27:222/monster/vidgod.git && cd vidgod
./setup/setup_venvs.sh        # venvs, SAM2 + MatAnyone clones, checkpoints, patches (~10 min)
./setup/smoke_test.sh         # fast lanes; add --roto for the full cutout pipeline

Everything heavy (venvs/, tools/, models/, library/, work/, tests/) lives untracked inside the working copy; the repo is just the scripts. Long jobs heartbeat to ~/.jobs/vidgod-*.status.

Typical session

bin/vg-transcode ~/old-tv-rips                      # → ~/old-tv-rips-prores
bin/vg-index ~/old-tv-rips-prores                   # overnight for a big archive
bin/vg-find "radical" --cut ~/mv/pulls              # grab every shot where someone says it
bin/vg-beats ~/mv/song.mp3 --fps 25                 # markers for the Resolve timeline
bin/vg-roto ~/mv/pulls/003_*.mov --grab-frame       # find your click coords
bin/vg-roto ~/mv/pulls/003_*.mov --point 970,220    # → ProRes 4444 alpha cutout

Notes & gotchas

• MatAnyone patch: torchvision ≥ 0.23 removed torchvision.io.read_video; setup applies patches/matanyone-cv2-reader.patch to its clone automatically. If you re-clone tools/MatAnyone by hand, re-run setup/setup_venvs.sh.
• MatAnyone's checkpoint comes from its GitHub release (the HF repo doesn't carry the file at the root path); its inference script insists on pretrained_models/ inside its repo dir, so setup symlinks it there.
• MatAnyone propagates forward from a first-frame mask — vg-roto --frame N trims the clip at N first. Prompt on a frame where the target is clearly visible.
• Scripts set /opt/homebrew/bin in PATH themselves (non-interactive SSH doesn't) and PYTORCH_ENABLE_MPS_FALLBACK=1 for stray unsupported MPS kernels.
• Rough speed on an M1 Ultra: MatAnyone ≈ 3.4 fr/s at 1920×804; whisper large-v3-turbo transcribes far faster than realtime; ProRes transcodes are ffmpeg-bound.

Phase 2 (not built yet)

ProPainter (object/logo removal), RIFE interpolation, Cutie interactive segmentation GUI, MODELBEAST farm operators for roto/index jobs, model mirror to NAS modelzoo. Upscaling old footage already exists fleet-side (seedvr2_upscale on the farm); Resolve's Optical Flow retime covers slow-mo.