Harvey Xia
Back to projects

imagePicker

Local-first photo workflowIn development

A more considered way to get from a full photo library to a useful selection.

imagePicker current group comparison with Qwen-generated family images
Group comparison · Actual backend, Qwen-generated images

The workstation brings photo grouping, explainable scoring and editing into one local-first workflow. Its assistant is designed to show a plan before making changes.

Status
In development, pre-release
AI in the product
Scoring, grouping and natural-language search
Evidence
Screenshots from the actual backend; source code
Limits
Not yet evaluated on real camera photos

The product question

How can a photographer review suggestions without losing control of the original files?

The design response

Keep selection separate from deletion. Make scores inspectable, group similar frames and require a review step before assistant actions.

What still needs proving

This is a pre-release project. These screenshots show actual backend results on synthetic images. They do not establish accuracy on real camera photos; broader scoring and end-to-end model evaluation remain incomplete.

imagePicker current grid with backend ratings of Qwen-generated images
Inspect the suggested scores.
imagePicker current natural-language search results for a cat query
Inspect a natural-language search.

Captured 30 October 2026 from the current WebUI and Rust backend using a library of 1,000 Qwen-generated images.

Explore imagePicker

Dated implementation evidence

  1. Only sharp faces start a new person group; blurred faces can join an established group.

    View commit · 6070f919

  2. Resolve yearless and relative dates against the session capture dates.

    View commit · 6cbd3902

  3. Document the current WebUI using a 1,000-image Qwen-generated library.

    View commit · d9c0d370

Verification and limits

The screenshots and dated commits support the workflow changes. Synthetic images are not a real-camera accuracy benchmark.

AI collaboration: attribution boundary

The records below show implementation and review outcomes. They do not record a complete prompt history or identify human versus AI authorship of individual changes. Commit history alone cannot establish that split; no percentage of authorship is claimed.