Pluck 1.0.0

Pluck 1.0.0 — the first release. Lift subjects out of photos, entirely on your Mac.

• 100% offline. Matting runs on-device — Apple Vision built in, or optional MIT-licensed BiRefNet models (83 MB, downloaded on demand, SHA256-verified against the signed app). No account, no upload, no telemetry.
• Three engines that disagree usefully. Vision is instant and honest about "no subject"; Clean Cut nails hard edges and line art; Fine Edges keeps hair and renders glass truly transparent. Switch per photo, compare live — measured, not marketed.
• A real Mac app. One window: drag in or ⌘V, grid or list, a before/after wipe in the inspector, batch export. English & 简体中文.
• Built for AI agents. The same engine ships as a pluck CLI — NDJSON output, semantic exit codes, never a GUI — plus an agent skill for Claude.
• Free forever. MIT. No watermark, no trial, no subscription.

Install: download Pluck.zip, unzip, drag to Applications. Signed & notarized — it opens without warnings. Requires macOS 26+ (Apple silicon).
CLI: pluck-cli-1.0.0-macos-arm64.zip, or build from source.

---

Pluck 1.0.0 — 首个正式版。把主体从照片里拎出来，全程在你的 Mac 上完成。

• 100% 离线。 抠图全部在设备端运行——内建 Apple Vision，可选 MIT 协议的 BiRefNet 模型（83 MB 按需下载，对照签名应用内的 SHA256 校验）。无账号、无上传、无遥测。
• 三个引擎，三种判断。 Vision 秒出结果、对"没有主体"诚实说不；利落边擅长硬边与线稿；柔细边保发丝、把玻璃抠成真透明。逐图切换、实时对比——实测结论，不是营销词。
• 真正的 Mac 应用。 一个窗口：拖入或 ⌘V，网格或列表，检查器里前后对比，批量导出。支持英文与简体中文。
• 为 AI agent 而生。 同一引擎以 pluck 命令行发布——NDJSON 输出、语义化 exit code、绝不弹窗——并附 Claude 可用的 agent skill。
• 免费，永远免费。 MIT 协议。无水印、无试用期、无订阅。

安装：下载 Pluck.zip，解压后拖入「应用程序」。已签名并公证，打开无任何警告。需要 macOS 26+（Apple 芯片）。
命令行：pluck-cli-1.0.0-macos-arm64.zip，或从源码构建。

Model weights v1 — BiRefNet_lite family (Core ML)

Core ML conversions of the BiRefNet_lite background-removal models, downloaded on demand by Pluck (never bundled with the app).

Asset	Model	Size	SHA256
BiRefNetLite.mlpackage.zip	BiRefNet_lite	82.6 MB	4614141d1e3ce84bd5713b995beabbb225e6e568522b296c4c7f48279228e817
BiRefNetLiteMatting.mlpackage.zip	BiRefNet_lite-matting	82.5 MB	7ea2236f7494ab7ac46b070847c00768cfd940249af62dd6ac1a367ed222f22a

Licence: MIT, Copyright (c) 2024 ZhengPeng (BiRefNet). Each archive carries the upstream LICENSE and a NOTICE with full attribution.

Provenance: converted from the original fp32 safetensors by Scripts/convert-birefnet.py (fp16, ImageNet normalization folded into the input layer, deform_conv2d rewritten in plain tensor ops). Numerical parity against the fp32 originals — MAE ≤ 0.0005, binary-mask agreement ≥ 99.96% on real photos — is recorded in docs/research.md appendix A.6.

Pluck verifies the pinned SHA256 of every download before installing (see product-plan §4.8); these hashes are compiled into the app's manifest.