Face & Expression Detection
Employs YuNet and MediaPipe keypoints to analyze faces, detecting closed eyes, squinting, head turn angles, and overall expression quality to cull bad shots.
YuNet ONNXMediaPipeGaze & Blink tracking
Core Features & Tech Specifications
Pixel Cafe combines a local-first AI processing engine with robust cloud collaboration and camera preset simulations.
AI Culling Pipeline
Our 5-stage pipeline runs on your GPU/CPU to filter duplicates, blurs, and bad expressions local-first.
Burst Grouping & Clustering
Groups similar bursts automatically within a customizable time window (e.g. ≤3s) and HSV histogram correlation, allowing you to select only the best frame per series.
DBSCANHSV Color HistogramsBurst reduction
Blur & Exposure Screening
Evaluates sharpness using Laplacian edge-detection gradients and monitors over/underexposure via brightness histograms to reject technical mistakes in stage one.
OpenCVLaplacian scoringExposure Histograms
Large Vision-Language Model Curation
An opt-in VLM scoring step (supporting Gemini, Claude, and ZhipuAI) that evaluates aesthetics, identifies main subjects/distractions, and scores photos from 1-10 with full reasoning.
VLM APIsAesthetic RatingMulti-lingual Reasoning
Personal Taste Learning
Builds a local pairwise preference model. As you star or reject images, the system learns your stylistic preferences and re-ranks bursts accordingly.
Pairwise LearningOn-device TrainingReversible Weights
WeChat Co-Culling & Sharing
Photographers sync projects to the cloud, allowing clients to select favorites instantly via WeChat.
Zero-Signup Client Mini Program
Clients scan a QR code or paste a share link in WeChat on iOS or Android, launching a mobile viewer to select photos instantly without registering.
WeChat EcosystemZero Sign-inMobile-first Webview
Real-time Selection Sync
Client choices on their phones are synced instantly to the photographer's desktop workstation via secure WebSocket cloud routing, eliminating email lists.
WebSocketsCloud synchronizationMulti-client review
Selection Limits & Timers
Photographers can enforce select caps (e.g. choose 50 out of 500) and set link expiration dates to keep project timelines and post-processing work on track.
Usage ControlSelection LimitsExpiry Dates
Retouching & Color Simulation
Local enhancement and camera profile mapping directly on your catalog.
Local AI Skin Retouching
A PyTorch-based UNet skin retouching model that runs entirely on-device, detecting faces and performing skin smoothing and tone leveling in batches.
UNet MaskingBatch Skin RetouchOn-device PyTorch
Nikon Picture Control Emulation
Import Nikon custom picture control profiles (.ncp curves/LUTs) to accurately simulate in-camera color curves, saturation, contrast, and sharpening.
NCP Curve Parse8-bit LUT mappingPreset Simulation
Developer & Production Tools
Built for robust deployments, packaging pipelines, and easy troubleshooting.
Breakpoint Checkpointing
Each stage of the AI pipeline checkpoints progress to a local SQLite database, allowing you to resume long-running tasks immediately after computer restarts or app crashes.
SQLitePeewee ORMCrash Recovery
Desktop Diagnostics Portal
Activate with a CLI parameter (--debug) or environment variables to open a dedicated logs panel to monitor Uvicorn, FastAPI, and model inferences in real-time.
Live Logs streamDebug FlagSystem Diagnostics
Smart Network Proxy Routing
A unified proxy policy with Force-Direct and Fallback-Direct options ensures reliable model index downloads from Cloudflare R2 and API calls through system proxies.
Proxy FallbacksOffline fallbackCloudflare R2
Cross-Platform Installer Packaging
GitHub Actions workflows compile signed, ad-hoc macOS DMG and Windows EXE packages, optimizing application bundles to reduce installer sizes.
PyInstallerInno SetupGitHub Actions CI/CD