Video Color Grade

Video Color Grade (assess → correct → looks → choose → LUT + apply)

Write the grade from scratch as code, generate several distinct named looks on the real footage, select one through explicit human choice or delegated agent judgment, then bake a portable .cube LUT and apply it to the full clip. No preset packs.

Core idea — correct, then style. Every look is corrective base + creative layer:

• corrective base makes the footage honest — for flat LOG/HLG it's the log→Rec.709 conversion; for already-Rec.709 footage it's WB neutralize + exposure + mild contrast. Every look shares this base, so "neutral" stays meaningful and looks differ by intent.
• creative layer is the look stacked on top (warm / cool / teal-orange / punchy / faded…).

Skin is the quality bar. On people — especially darker skin — teal_orange and cool_desat routinely throw a green/grey cast. Judge every look on faces first; keep midtone warmth, push the tint into shadows/highlights/background, prefer vibrance over raw saturation. Don't push the white balance all the way to neutral or skin goes lifeless.

Requirements: ffmpeg/ffprobe on PATH, Python with numpy + Pillow.

Contract

• Every look = base + creative layer; the base is shared by all looks.
• All looks render on the same representative frame (a fair, controlled comparison), labeled, at full resolution.
• The full apply uses -c:a copy (audio never re-encoded → A/V sync identical) and keeps source duration, fps, dimensions.
• Complete look selection before delivery. Use human mode when the user asked to choose; use agent mode without pausing when the user delegated the choice or requested an autonomous run. In both modes, preserve the rationale in the durable plan.

Project protocol workflow

Store the durable decision in work/color-grade/grade-plan.json:

{
  "schema_version": 1,
  "target": "base-video",
  "base": "eq=contrast=1.05",
  "looks": [{"name": "clean", "chain": "null"}],
  "selected_look": "clean",
  "selection_mode": "agent",
  "selection_rationale": "Neutral correction preserves natural skin and source lighting.",
  "selected_lut": "../../final/selected-color-look.cube",
  "evidence_refs": ["media:source"]
}

Use base-video by default so content cards and captions are not recolored. composite is valid only when intentionally grading already-composited pixels. Read shared work/understand/media.json and visual evidence. Generate review/02-color-grade/choose-color-look.jpg and review/02-color-grade/skin-tone-check.jpg; optionally generate the short selected-look-preview.mp4. Record the decision and evidence in review/02-color-grade/selected-look.md.

Set selection_mode to human after an explicit user choice. Set it to agent when the user delegates the decision or requests an uninterrupted workflow, and choose conservatively with skin tone, highlight retention, and neutral balance as the priorities. Always write a non-empty selection_rationale; do not pause in agent mode.

Bake the chosen LUT to final/selected-color-look.cube. When that LUT is shipped, set selected_lut in the grade plan so the shared renderer consumes that exact file. A baked full-look LUT already contains the corrective base; the renderer must apply lut3d alone, not base + lut3d.

Record the operation's exact input revisions in based_on, increment its integer revision when the plan or selected input changes, set the operation check to pass only after the review files are complete, and contribute the full protocol fields:

{
"target": {"sequence": "main", "scope": "base-video"},
"effects": {
  "changes_timeline": false,
  "changes_geometry": false,
  "changes_video_pixels": true,
  "changes_audio": false,
  "adds_track": null
},
"check": {
  "status": "pass",
  "report": "../review/02-color-grade/selected-look.md"
},
"render": {
  "kind": "video-filter",
  "target": "base-video",
  "plan": "color-grade/grade-plan.json"
}
}

The shared renderer validates that selected_look exists, applies the grade after timeline transforms and before overlays, and encodes the final video once:

python skills/video-understand/scripts/build_render_plan.py .
python skills/video-understand/scripts/render_project.py work/render/render-plan.json

Standalone compatibility inputs

• The source video.
• A looks.json (base + named looks); start from looks.example.json and retune the base from what assess.py reports.

All existing scripts continue to accept legacy looks.json. gradelib.load_spec() also accepts canonical grade-plan.json, so contact-sheet, clip, LUT, and standalone apply commands remain compatible during migration.

looks.json schema

{
  "base": "<corrective base: log->Rec.709 conversion, OR WB+exposure+contrast for normal footage>",
  "looks": [
    { "name": "clean_neutral", "label": "1 · CLEAN NEUTRAL", "desc": "corrected only", "chain": "eq=contrast=1.02" },
    { "name": "teal_orange",   "label": "4 · TEAL & ORANGE", "desc": "cinematic",      "chain": "colorbalance=...,eq=..." }
    // a look's full ffmpeg filter = base + "," + chain. Set "prepend_base": false to use chain standalone.
  ]
}

Workflow

1. Assess — never grade blind.

   python scripts/assess.py SOURCE.mp4 --out work/
   

   Reports log-vs-Rec.709, white-balance cast (signalstats U/V vs 128), exposure (Y), picks a representative frame (work/assess/frame.png), and suggests base tweaks. Retune looks.json base accordingly (e.g. warm cast → add a little blue; log → put the log→709 conversion in base).

   • The rep frame is picked by median luma, not by faces. On a single-shot talking-head that usually lands on a good face frame, but on B-roll-heavy / multi-shot footage it can pick a faceless frame — and then face_skin_check.png is useless (skin is the quality bar). Force a known face moment with --frame-time SECONDS.
   • The WB/exposure stats come from one ~10s signalstats window at 10% of duration — fine for constant light; for footage whose light changes, move/lengthen it with --win-start SECONDS / --win-dur SECONDS.

2. Render the looks for choosing.

   python scripts/render_looks.py work/assess/frame.png looks.json --out work/looks
   

   Produces work/looks/looks_compare.png (original + every look, labeled, full-res) and face_skin_check.png (zoomed skin strip). LOOK at both. Skin is the bar — retune or drop any look that greys/greens skin, then re-render. Off-center subject? add --face-crop W:H:X:Y. Treat those PNGs as working intermediates; publish the approved review copies as review/02-color-grade/choose-color-look.jpg and review/02-color-grade/skin-tone-check.jpg.

3. (optional) Motion previews of the top 2–3 so they're seen moving (audio kept):

   python scripts/make_clips.py SOURCE.mp4 looks.json --out work/clips \
     --pick clean_neutral,warm_filmic,teal_orange --with-original --ss 148 --t 8
   

4. Select one look. Present the contact sheet and wait only when the user asked to choose. Otherwise select it in agent mode and continue without a decision pause. Write selected_look, selection_mode, and selection_rationale before delivery.

5. After the pick — bake the LUT and apply to the full clip.

   python scripts/bake_lut.py looks.json --name CHOSEN \
     --out final/selected-color-look.cube --verify-frame work/assess/frame.png
   python scripts/apply_grade.py SOURCE.mp4 --looks looks.json --name CHOSEN --out out/graded.mp4
   #   or apply the delivered LUT: --lut final/selected-color-look.cube
   

   apply_grade.py verifies duration/audio are kept, prints the WB shift (U/V toward 128), and drops a spot frame to eyeball.

   • The apply re-encodes the video once (lossy); audio is -c:a copy. Quality/time are set by --crf (default 18) and --preset (default slow). slow is fine for small/ short clips, but on 1080p/4K hour-long footage it's a large, silent time cost — drop to --preset medium for delivery (or veryfast for a quick preview), slow only for final.
   • Match the delivered video to the delivered LUT: the chain (--name) and the baked 33³ .cube (--lut) can differ by up to ~5/255 on steep curves (bake_lut.py reports this max|d|). Invisible for most work, but if you're shipping the .cube and want the rendered video to match it exactly, apply with --lut final/selected-color-look.cube, not --name.

Design notes / gotchas

• Grade the CLEAN cut, then composite overlays on top of the graded footage. If you must grade a video that already has burned-in captions/cards, do NOT trust assess.py: its WB/ exposure stats and frame pick are skewed by the graphics (dark top/bottom scrims drag Y down, bright caption text / teal accents skew U/V, and the frame pick can land on a caption-heavy frame) — you'd be correcting the overlays, not the footage. Reuse a base tuned on the clean footage, and expect the grade to tint the overlays, so use only a subtle/corrective look (clean_neutral); teal_orange/vintage_faded will visibly recolor the text/accents.
• One look/LUT covers all related clips. Once you've picked a look, reuse the same looks.json / .cube across every related clip (e.g. clean cut + the overlaid version) — do NOT re-run the two-phase pick per clip.
• Correct, then style. The base is the generalization of the deck's S-Log3→Rec.709 step. For normal footage the base is WB+exposure+contrast; for log footage put the conversion there.
• Skin first. teal_orange/cool_desat are the usual skin offenders on darker subjects — keep midtones warm, tint shadows/highlights/bg, use vibrance. Don't neutralize U/V fully.
• Labels are PIL PNGs overlaid by ffmpeg — never drawtext. On Windows the drawtext fontfile path (drive colon) is unreliable. Fonts are set at the top of gradelib.py.
• Drive-colon paths break ffmpeg filtergraph options that take a path (lut3d, metadata=print:file=). The scripts run ffmpeg with cwd = the file's folder and reference it by basename. Keep this pattern for any new path-taking filter.
• Outputs land in the --out dirs you pass (durable), not a system temp dir.
• The face/skin strip uses a fixed centered upper-middle crop, not face detection. If the subject isn't roughly centered, pass --face-crop W:H:X:Y to render_looks.py or the skin strip will show the wrong region.
• apply_grade.py's verify spot frame is at 20% of duration — a different frame than the one the looks were chosen on. If that 20% point is faceless, eyeball skin on another frame instead (or pass --frame-time to assess.py so the chosen frame is a known face moment).
• The LUT bake is exact, because every filter here is a per-pixel point op: bake_lut.py pushes a 33³ identity grid through the real chain and self-checks against it (mean|d| should be well under 1/255; a few-level max|d| is just interpolation).
• Selection before delivery, always. Human choice and delegated agent choice are both valid, but the plan must record which occurred and why.