AI Photo: Design and Technical Documentation of a Browser-Based AI Photo Studio Combining On-Device Computer Vision with Optional Generative Image Models

Photo-editing tools that use artificial intelligence (AI) typically upload the user's images to a cloud service, which raises privacy, cost, and provenance concerns, and which blurs the distinction between ordinary editing and the generation of synthetic content. This paper describes the design and technical implementation of AI Photo, a free browser-based photo studio published by AI Verse. The application draws an explicit boundary between operations that run entirely on the user's device and operations that require a generative cloud or local model. On-device operations include tone and color adjustment (15 parameters), 12 filter presets plus an automatic enhancement, background removal and replacement with neural segmentation models, object removal with classical inpainting, noise reduction, local-contrast lighting correction, neural 2× upscaling, cropping to nine aspect ratios, optical character recognition in 12 languages, face-aware portrait and passport-style framing, social-media graphic composition, and a photo library with automatic tagging and natural-language search using image–text embeddings. Generative operations—text-to-image creation and prompt-based content edits such as “add a sunset sky”—are routed to a provider the user selects (a free hosted service, Pollinations, OpenAI, Google Gemini, Hugging Face, or a local Stable Diffusion server). A language-model planner translates plain-language editing requests into a sequence of editing steps and is instructed to prefer local steps over generative ones. Images and library data are stored in the browser, and an image is sent to a provider only when a cloud feature is used. The paper documents the architecture, models, features, and a step-by-step tutorial, and discusses limitations, licensing, and safeguards against deceptive use. No formal evaluation of output quality, segmentation accuracy, search relevance, usability, or user behavior has been conducted; the benefits described are proposed rather than demonstrated.