PixCull

AI photo culling for working photographers. Six calibrated scoring axes · burst folding · style-aware personalization · Lightroom / Capture One round-trip. Judged by MiniMax M3, with the local measurements handed to it as evidence — or fully on-device with --vlm-mode off.

AI 摄影分拣:6 轴评分 · 连拍折叠 · 个性化学习 · Lr/C1 双向 round-trip。 MiniMax M3 云端判图,本机实测指标作为证据一并送入;--vlm-mode off 可全程不出网。

Install

pip install pixcull

Faces, closed eyes and face blur are checked with MediaPipe, an optional extra: pip install 'pixcull[face]'. Without it a run says so and the other checks are unchanged.

Python 3.11–3.13. First run downloads the optional scoring models to ~/.pixcull/models/; the measurements they make run on-device and are Apple-silicon accelerated. Whether the judge is local or cloud is a separate question — see the last two bullets below.

Upgrading from 3.53.1? The dependency floor moved to numpy 2 and torch 2.7, and five other things behave differently. Read the upgrade notes first.

Quickstart

# score a folder of photos (JPG / RAW) — keep/maybe/cull + per-axis rubric
pixcull run /path/to/photos -o ./out

# write the decisions back as XMP sidecars for Lightroom / Capture One
# (sidecars land next to each original; -t collected gathers them instead)
pixcull export ./out

# score a video → temporal windows + reel candidates
pixcull video clip.mp4 -o ./out

# fold near-duplicates, build a contact sheet, learn your taste, …
pixcull --help

# open the review workspace in your browser (v2.31 — now packaged)
pixcull serve

The pip package ships the full scoring engine, the CLI and the interactive review workspace: keyboard-first grid, ⌘K palette, per-axis "why" explanations, n-way compare, maybe-resolution queue, XMP export. pixcull serve stores runs under ~/.pixcull/runs by default (--root to change, --host 0.0.0.0 to share on a trusted LAN).

Highlights

• Glass-box scoring — every keep/cull carries a per-axis breakdown (technical / subject / composition / light / moment / aesthetic) and a plain-language "why", not just a number.
• Learns your taste — corrections feed a personal profile that tilts the axis weights toward what you demonstrably value.
• Burst & near-dup folding — stacks collapse to the peak frame with a one-key compare.
• Video too — temporal scoring and reel-candidate detection with the same glass-box treatment. With the optional audio model (pixcull models pull audio-tagger, 16 MB) it also marks laughter, applause and music on the timeline; without it, it says it did not listen.
• 13 UI languages, dark/light studio-neutral themes, WCAG-conscious.
• You choose where it runs — with a MiniMax key configured, cloud judging is the default and your photos are uploaded to MiniMax (it asks once, and declining keeps you on-device); with no key, a bare pixcull run never leaves the machine. --vlm-mode off forces the on-device path outright. No telemetry either way.
• Cloud judging needs a funded MiniMax account. It is their API and their bill, paid with your own key. Without balance the API returns 402 and PixCull falls back to the on-device rule stack, so nothing breaks; it just stops being the thing the first bullet describes. pixcull m3 doctor tells you which of those you are in, and distinguishes 'no balance' from 'bad key' — they look identical otherwise.

Links

• GitHub + full README
• Releases
• ModelScope mirror

MIT © Chris Chen