CommunityImagegithub.com

shootpack

Process a real-estate shoot folder (DJI drone and/or camera such as the Sony A1 II, mixed is fine) into client deliverables in one step — HDR photos from the raw brackets (LightCraft), culled picks/alternates, and a 60 s Rec.709-graded video. Use when the user points at a shoot folder ("process this shoot", "package Maginot", "run shootpack on ~/Desktop/X").

What is shootpack?

shootpack is a Claude Code agent skill that process a real-estate shoot folder (DJI drone and/or camera such as the Sony A1 II, mixed is fine) into client deliverables in one step — HDR photos from the raw brackets (LightCraft), culled picks/alternates, and a 60 s Rec.709-graded video. Use when the user points at a shoot folder ("process this shoot", "package Maginot", "run shootpack on ~/Desktop/X").

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/mattgiss/shootpack/tree/main/skill

Installed? Explore more Image skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

What does shootpack do?

Pipeline lives in ~/shootpack (git repo). Entry point: shootpack <folder> [--ceiling M] [--picks N] [--length S] [--log CURVE] [--latlon LAT,LON] [--plan]. Start new shoots with --plan (inventory, cull and edit plan, no exports).

Devices

Every file's camera is detected (EXIF Make/Model, video encoder/model tags, sidecars) and read with its profile in devices/<id>.json; unknown cameras use devices/generic.json. The first log line lists what was found (devices: DJI Mavic 4 Pro (200 photos, 8 clips), Sony A1 II (...)) — check it.

  • DJI Mavic 4 Pro (dji_mavic4pro): XMP drone geometry → sun detection, framing-aware cull, ceiling flags, altitude-aware video; SRT telemetry, LRF proxies.
  • Sony A1 II (sony_a1ii, EXIF Model ILCE-1M2): ARW brackets (+ JPG/HEIF pairs, also across cards) grouped by exposure bias + drive mode; clip XML sidecars (C0001M01.XML) give start time, frame rate, gamma and gamut; SUB/ proxies. No camera geometry: no auto sun profile (set "sun" in cull.json by hand), the cull goes by image content (near-identical frames = duplicates) and names shots by lens (24mm, 70mm-portrait).
  • Fujifilm X-E4 (fujifilm_xe4, EXIF Model X-E4 on RAF, JPG and MP4): single frames, RAF + JPG pairs. The JPG is the recipe render (Classic Chrome etc.; the recipe is in the RAF's EXIF, see photo.recipe_tags); LightCraft's own RAF develop is dark and flat and ignores it, so fit a develop profile to the camera JPEG before exporting. Clips in the same DCIM folder, no sidecars; film-simulation clips are Rec.709 (look only, no technical LUT), F-Log is detected from the signal. Offload: shootpack offload <card> --device XE4 (drop folder Drops/XE4/). Develop: recipe_fit.py fit LUT.npz RAF=JPG,... on the shoot's own pairs (pair RAF and JPG by capture time: the numbers can differ by one), then fuji_develop.py batch LUT.npz cull.json <DCIM dir> Deliverables/Photos (cull.json keys are the JPEG number -> {raf, jpg, name}). The camera clock may be on home time (one wedding shoot: -06:00, local = +8 h). Event clips with live sound: clip_edit.py edits/<spec>.json Video/<name>_4K.mp4 (no music, L-cut audio; stabilisation stays off unless motion.py proves it helps). See agent-docs/photos/fuji-recipe.md.
  • DJI Osmo 360 II (dji_osmo360ii, .OSV found by the MP4 Encoder tag Osmo 360 II): two fisheye streams, so its clips are set aside (the log says "N 360 clip(s) ... render them with osmo360.py") and never reach the real-estate edit. osmo360.py CLIP.OSV Deliverables/Video writes a graded flat reframe (keyframed move) and a graded equirect 360. The stitch is uncalibrated and the horizon is not levelled; the lenses are gain-matched in the overlap. Colour mode is read from the signal inside the image circles (Normal colour vs D-Log M, verified on real clips of both). D-Log M gets a stand-in LUT fitted to the camera's own proxy (luts/interim_dlogm_to_709_osmo360ii.cube) until DJI's Osmo 360 D-Log M LUT is in luts/official/. --spec edits/osmo_car_v1.json renders a keyframed excerpt of a long clip (a tight, exposure-lifted crop of the driver in the car test). A flat crop has 18.6 x h_fov real pixels across: 4K of a tight crop is upscaled (agent-docs/video/osmo360.md). The camera has internal storage and an SD card (two volumes). Offload: shootpack offload <card> finds it itself (Drops/Osmo360/).
  • Dual-slot recording: identical files on both cards are detected (name + capture time + size, then hash) and the slot-1 copy is used; the log says how many were skipped.
  • No GPS anywhere (camera-only shoot): sessions are named by local clock time (Morning < 10:00, Midday < 15:00, Afternoon < 17:30, Evening), and the grade follows the name (Midday/Afternoon: midday look; else golden). Pass --latlon to name them by the sun instead. Sony shots have no geometry, so sun shots get no automatic sun profile: set "sun": {"G18": "A"} in cull.json by hand (auto develop lifts a dark sun-through-glass frame flat).
  • A new camera: copy the closest profile, change detect.models, sensor/FOV, bracket rule and log curves.

Steps

  1. Run it. shootpack "<folder>" --ceiling <user's LAANC/airspace ceiling, default 120> in the background (photos ≈ 25 s per bracket, video ≈ 4 min). It caches everything in <folder>/Deliverables/_shootpack/.
  2. Review the cull visually — open Deliverables/_shootpack/<session>/all_stations.jpg and each Deliverables/Photos/<session>/Picks_contact_sheet.jpg. The auto-cull can't see content, so check for and fix:
    • the house cropped at a frame edge → move to rejects
    • a person (often the pilot in hi-vis) in frame → alternates, note "PERSON-IN-FRAME"
    • near-duplicates / boring repeats → alternates; strong shots left in alternates → picks
    • shots over the ceiling are already flagged in alternates — mention them, don't hide them Edit <session>/cull.json (keys G01…; values are short names), delete the affected outputs, re-run with --only photos. Sun shots are detected automatically (sun inside the frame by solar geometry + still clipped in the darkest bracket) and developed with a sun profile: A = sun is the subject (-1.9 stops vs auto), C = house is the subject (-0.6 stops, house stays bright). Override per shot in cull.json: "sun": {"G21": "A", "G07": "off"}. Starbursts: if a bracket frame shows crisp diffraction rays around a visible sun disc (look at the sun in each frame), add "starburst": {"G21": "0209"} (frame number) and re-run --only photos --redo; never for suns behind cloud (those rays are sunbeams and already survive the merge). Check sun shots for pink clipped clouds.
  3. Review the video — open Deliverables/Video/*_storyboard.jpg (one frame per 3 s). Replace shots that are takeoff/landing, the pilot, too dark, or repetitive by editing Deliverables/_shootpack/edit.json (clip, start, len, speed, exposure in stops) and re-run --only video. Use the clips' proxies (DJI .LRF, Sony SUB/) and DJI .SRT (rel_alt, position) to find better moments. Camera clips have no telemetry, so takeoff/landing trimming and altitude bands don't apply; 24p/30p clips are never slowed down (only 50p+ can feed slow motion).
  4. Grade — with DJI's official LUT in ~/shootpack/luts/official/ (Maginot: "DJI Mavic 4 Pro D-Log to Rec.709 V1.cube"), every shot gets DJI's LUT as the technical base plus our creative look on top: luts/looks/ dlog_golden_* (golden hour: warmer, calmer blue, opened shadows) or dlog_midday_* (more contrast/punch). luts/looks/make_looks.sh rebuilds them (build_look.py: _input.cube = official LUT with the exposure applied in log, for FilmCraft's Input slot; _look.cube = Rec.709→Rec.709 for the Creative slot; _composed.cube = both, for ffmpeg). Put exposure lifts BEFORE the official LUT (--pre-exposure, stops in log): DJI's LUT is darker than its own JPEGs, and lifting after it clips skies (v5 of Maginot came out 30-50% too dark because of that). Without an official LUT the curve-based warm_print_v6 / warm_print_midday_v2 are used. Always verify a new official LUT matches the footage's profile (see below).
  5. Music (trailer) — Epidemic Sound through the user's OAuth connector (never an API key). Search instrumental "build" tracks whose bar length matches the montage cuts (75 BPM = 3.2 s bars) and read each waveform's energy curve: you want a quiet intro, a long build and one big drop. DownloadRecording (WAV, FULL) into Deliverables/Music/, then music.analyze(wav, given_bpm): check that the drop lands on a downbeat (the grid's on/off-beat phase comes from the kick/bass, the bar phase from the biggest low-end jump). Pick jumps X→Y where bar_chroma similarity of bar Y to bar X+1 is >= 0.9 (same chords, no audible splice). Time the picture in whole frames on the bar grid (template: edits/make_trailer_v7.py): cuts on downbeats, finale on the drop, end card on the release. trailer.py muxes spec["music"] into every output; --music-only re-edits and re-muxes without re-rendering; --size 1920x1080 makes a 5-minute timing check (measure cuts against beats before the 6K render). Client taste (reference shoot, v8-v11): full 16:9 with no letterbox ("letterbox": null), no text cards, no dissolves or dips to black, hard cuts on the beat, end on the last shot (a 1 s "end_fade" under the music's tail is fine). Smooth shots with "stabilize": a number = Gaussian sigma in s (1-2), {"poly": 2} = an even move. Without the letterbox every vertical correction costs zoom, so keep sigma moderate (the reference shoot's finale: 1.25 = 6.6% zoom, 2.0 would be 12%); "reframe_y" only helps with a letterbox. To keep more of the house at the end of a move, slow the shot so it ends earlier in the move. Measure shake on the render (stabilize.track on the preview) before and after; a drone's own stop-and-reverse can only be softened (heavier sigma + 0.85x), not removed. Epidemic's EditRecording returned no edit whenever requiredRegionsAtOffsets was set, and its unconstrained 60 s edit dropped the drop entirely (Oct 2026), so we cut the track ourselves. Client work needs a commercial Epidemic plan; ask.
  6. Deliver — send the picks contact sheets, the 720p preview (<30 MB) and storyboard with SendUserFile; summarize from Deliverables/SUMMARY.md (counts, flags, what's interim). Commit any reusable code changes to the repo.

Show progress

For every change, give a before/after: update review_site/<shoot>/progress/*/manifest.json (grade / photos / detail / cut / music) and rebuild with progress_page.py (progress_music.py draws the music-fit chart); send a phone-friendly sheet or stacked comparison video with SendUserFile.

Facts that matter

  • Always develop from the raws (DJI DNG, Sony ARW); JPGs are only metadata/reference. ARW-only shots get a reference JPEG extracted from the raw into _shootpack/previews/. LightCraft renders ARW with a look fitted to Sony's own JPEG (not the muted neutral fallback) but has no Sony lens profiles: distortion/vignetting stay as shot. Its HDR merge of ARWs writes a DNG with no colour matrix or white balance (develops green): the pipeline writes the camera's calibration in (ColorMatrix1 from LibRaw's Adobe-derived table via rawpy, AsShotNeutral from the shot) and develops that file, so HDR_DNG/ files from Sony brackets open with correct colour in any raw editor.
  • LightCraft writes *-HDR.dng next to its source; the pipeline moves them to Deliverables/HDR_DNG/. Never modify originals.
  • DJI MP4s are stored rotated (rotation=-90); ffprobe reports portrait dims but they play landscape.
  • Sony cards: copy whole DCIM/ + PRIVATE/ trees per slot (Finder can't mount a camera in MTP mode — use a card reader or USB Mass Storage); without the M01.XML sidecars the clips' gamma/gamut fall back to the signal alone.
  • SRT DiffTime: 16ms = 59.94p → half-speed conform to 29.97 for smooth orbits.
  • Log curves are decided per clip from the signal: the luma code-value distribution of ~10 frames (p95 = where the whites sit) is checked against each curve's band in the profile, then the camera's label (DJI SRT color_md, Sony XML CaptureGammaEquation), then --log. Order of trust: an explicit --log, then the signal, then the label. Every disagreement is a WARNING in the log and in SUMMARY.md — read it before rendering. D-Log vs D-Log M: in daylight D-Log keeps whites around 0.55 code (super-whites above for sky/sun; Maginot p95 0.44-0.69); D-Log M puts them near 0.8+. A D-Log M LUT on D-Log footage renders whites as ~50% grey (flat, dull) — Maginot was D-Log. Too dark (night) or between bands = inconclusive → the label decides. _shootpack/log/*.json caches the stats.
  • LUT matching per device/curve/gamut: official LUTs in luts/official/ by normalised file-name tokens ("D-Log M" counts as D-Log M; a Mavic 4 Pro LUT wins over other DJI LUTs). DJI without an official LUT: curve-built warm print grades. Sony S-Log3: FilmCraft's built-in builtin:slog3-sgamut3cine-to-rec709 (or -sgamut3-, from the XML's CaptureColorPrimaries), exported to luts/builtin/; our D-Log looks are re-composed on it as luts/looks/slog3_*_composed.cube, with the pre-exposure matched so 18% grey lands where DJI's input LUT puts it (mixed DJI + Sony edits match). S-Cinetone/Rec.709/HLG clips need no decode: the _look.cube alone.
  • Exposure in edit.json is lifted in code values per stop of the clip's curve (lift: D-Log 67, S-Log3 79).
  • LightCraft sun shots: keep light.whites at 0 and light.highlights at -60 or gentler; stronger recovery or a magenta tint drags clipped cloud below white and it renders pink.
  • 6K renders: trailer.py renders in sections (split at hard cuts and dips to black) — 14 parallel 6K decodes run out of memory on 16 GB.
  • Watch free disk: each HDR DNG ≈ 300 MB; the pipeline stops under 5 GB free.

Related Skills