Local-First
Local-first by default
The full curation pipeline runs on your machine. Cloud sync is optional—your originals never leave unless you choose to.
RAW · JPEG · HEICM-series · CUDAFree during beta
v0.4.34macOS · WindowsCull a wedding
Cull a wedding
in the time it takes to brew a coffee.
Pixel Cafe is the local-first AI curation workstation for working photographers. Sharpness, face, cluster and taste ranking—on your machine, before your coffee cools.
Also known as: Pixel Cafe · Pix Cafe · 像素咖啡 · PixCafeAI
60+frames analyzed / sec
5AI pipeline stages
92%best-shot retention
EN · 中 · 日languages
Adaptive ranking+18.4%Cluster hit-rate this week
AI 流程202 / 2025-stage analysis complete
Pending 151★ Starred 25✕ Rejected 26✦ Smart star
28.78★ 5/5
27.03★ 0/5
24.23★ 0/5
26.27★ 0/5
25.26★ 5/5
22.10★ 0/5
26.43★ 0/5
24.51★ 0/5
23.18★ 0/5
25.70★ 5/5
21.99★ 0/5
24.04★ 0/5
Built for working photographers
◐ Wedding Studio◆ Sports Pro▣ Portrait Atelier⌬ Commercial✦ Event Crew◇ Documentary
AI Pipeline
A 5-stage AI pipeline
After import, every frame flows through five stages. Each step focuses on one signal so results are traceable and reversible.
STAGE 01
Technical quality
Catch blur, motion shake and exposure problems before anything else.
SharpnessSTAGE 02
Face & expression
Extract facial features, expression and gaze for fast people-based filtering.
InsightFaceSTAGE 03
Burst clustering
Group bursts and near-duplicates so you never re-judge the same frame twice.
DBSCANSTAGE 04
Taste ranking
Pairwise learning updates same-cluster order live, using your star and reject history.
PairwiseSTAGE 05
VLM aesthetic
Optional high-token aesthetic scoring, runs as an opt-in step so import stays fast.
OptionalWhy Pixel Cafe
Curation isn't manual labour—it's accelerated judgment.
We didn't build another cloud library. We took the part that actually steals your hours—picking the best frame in each burst—and handed it to a local AI that keeps learning what you like.
Adaptive Ranking
Learns your taste over time
Every star and reject becomes a training sample. Pairwise learning re-ranks each burst against your evolving taste.
1,000 样本起步PairwiseReversible weights
Batch & Deliver
Import to delivery, one flow
Batch retouch, smart starring, scene-level rejection and shareable client links all live in the same workspace.
AI 批量美肤智能标星协同选片
Workspace
Desktop handles local AI work. Web handles collaboration and review.
Showing the two surfaces separately up front makes the product model clear: local culling on desktop, shared review on the web.
Desktop workstation
Built for photographers and retouchers who need import, analysis, and project data front and center.
Importlocal ingest
5 StageAI analysis
Liveproject metrics
- Import: bring in RAW or JPEG folders, generate previews, and build project structure locally.
- Analyze: run the full 5-stage AI pipeline and inspect screening, grouping, and ranking output.
- Data: desktop users should primarily see pending, starred, rejected, and training-progress data.
Web collaboration hub
Built for teammates and clients who mainly need cloud access, shared review, and collaboration data.
Cloudproject access
Multi-usershared culling
Sharereview links
- Collaborate: photographers, retouchers, and clients can review the same project together.
- Data: web users should primarily see shared status, online review state, collaborative results, and shortlist outcomes.
- Capability: the web experience emphasizes cloud review and collaboration rather than local ingest or heavy analysis.
Live workstation
The workstation reads like a selection assistant—at a glance, fully traceable.
From the moment you import, see pending / starred / rejected counts plus training sample progress. Every AI call carries its reasoning—rollback is one click away.
- ✓Global filter overview with live distribution
- ✓Every AI call is inspectable and overridable
- ✓Auto-enables personalisation once sample threshold is met
- ✓Shared client links for in-browser co-selection
Pending
151
Starred
25
Rejected
26
Personal training progress
Current 231 / target 1,00023.1%
In production
Tested in real shoots, not curated demos.
Wedding
Wedding studio
Run the whole event in one pass: rings, family portraits, toasts auto-clustered, near-duplicates collapsed, the best frame of each group surfaces on top.
Travel
Travel album
Repeats of the same landmark and burst variations cluster together. Blur and exposure issues land in reject queue, so what's left is the stories you'd actually tell.
Family
Family portraits
Time-sensitive moments—a baby's smile, a mother-child glance—surfaced by face plus expression scoring. One click of batch retouch keeps skin tones consistent across the set.
“
Pixel Cafe has completely transformed the way I select and deliver photos. What used to take hours now takes minutes.
FAQ
Questions before you download.
选片过程中需要把照片上传到云端吗？
不需要。Pixel Cafe 的核心选片流水线完全在本地运行——模糊筛选、人脸识别、连拍聚类、个性化排序都使用你机器的 GPU / CPU。云端同步只是可选项，用于协同选片或多设备工作时手动开启。
是否支持 RAW 文件？
支持。我们通过 exiftool 抽取预览图来加速 RAW 处理，覆盖 Sony、Canon、Nikon、Fujifilm、Sigma 等主流相机品牌，单文件也可以是 HEIC / JPEG / TIFF。
AI 流水线具体包含哪些阶段？
Stage 1 锐度（模糊与曝光）、Stage 2 人脸（InsightFace 特征）、Stage 3 聚类（连拍 / 相似构图）、Stage 4 个性化 Ranking（pairwise 学习）、Stage 5 VLM 美学评分（可选，按需运行）。
Pixel Cafe 会不会收集我的照片用于训练？
不会。我们不会上传、收集或用你的素材来训练模型。本地优先是核心承诺，云端同步只在你主动启用时才会传输被你选中的内容。
支持哪些操作系统与硬件？
macOS 12+（Apple Silicon 与 Intel）、Windows 10+（x64，可选 CUDA 加速）。Apple M1 起的机器上可享受 Metal 推理加速，常见婚礼项目（~3,000 张）通常在 4-6 分钟内完成全部 5 阶段分析。
个性化排序需要多少样本才会生效？
默认阈值是 1,000 个标星 / 淘汰样本。在到达阈值前，Pixel Cafe 会使用全局基础权重；达到后会自动启用个性化 Ranking，并允许你随时回退到基础权重。
Get started
Start local. Let the ranking learn your taste from there.
v0.4.34 is out: 5-stage analysis, batch retouch, and shareable client selection links.