Latent Vision Studio GitHub
Ecosystem

Seven repositories,
one job each.

Each module does one thing and can be read, replaced or left alone on its own. Below is what each one means for your photographs. If you are the sort of person who wants the database schema and the scheduler, it is a click away — folded down, out of the way of everyone else.

Latent Vision Studio
The card goes in; the thinking starts. Previews are on screen straight away, then captions, scores and edge maps are worked out for every frame while you get on with the culling.
latent_vision_studio
What is happening

Two ingest paths: sequential with kernel I/O telemetry on Linux, concurrent CPU-producer/GPU-consumer on Windows. Florence-2 dense captions, CLIP aesthetic scoring and HED soft-edge maps, committed to a resumable SQLite ingest database.

Selection Assist
The culling pass you already do, with a second opinion on screen. It rides over FastStone or digiKam, tells you what was noticed about the current frame, and takes it away on the number keys.
LVS_Select
What is happening

A frameless Qt6 overlay reads the ingest database; AutoHotkey v2 drives the viewer’s move dialog over a named pipe into select1–5 buckets. A tasker then verifies hashes, resolves the culled previews back to their RAWs, writes EXIF ratings with ExifTool and produces the handoff manifest.

Batonpass & Rifinire
You describe the edit in the chat you are already in — “warm skin 300K, keep the sky” — and it comes back as a real RawTherapee recipe with the masks to match, without moving a 120 MB file anywhere.
LVS_MCP
What is happening

An MCP server that compiles model intent into clamped, camera-aware settings and renders them locally through RawTherapee and ImageMagick. Batonpass handles global edits, Rifinire the mask path; SEQ mode pins both to the same frame so the edit and its mask cannot drift apart by one RAW.

Agentic Editing
Hand it a folder on the way to bed. Strong frames come back finished, doubtful ones come back flagged with the reason, and no frame is quietly guessed at.
LVS_Agentic
What is happening

As-You-See-Fit drives autonomous editing passes; Sperpero generates the masks they need. It is the least predictable module by nature — different foundation models behave differently, and some behave strangely — so it is built to be supervised rather than believed, with keys that can be kept off the machine entirely.

Sprezzatura
Local adjustments that land where they were aimed: on skin, on eyes, on the backdrop — with the gradient of a sky still intact when it is handed back.
LVS_Sprezzatura
What is happening

A command-line compositor that isolates colour-coded regions from the masks, applies its adjustment recipes in float32, blends in two passes (negative before positive) and writes layered PSD, ORA or XCF alongside a quality-capped JPEG and an audit log.

Mask Gallery
The masks shown next to the photograph they were made from, so you can see at a glance whether one slipped — and fix it by dragging rather than by reading a log.
LVS_mask_gallery
What is happening

Alignment by edge chamfer scoring, SSIM, aspect ratio and time consensus, with unmatched masks held in an orphan tray for manual assignment. Generated masks are pre-processed to strip embedded watermarks before they are used.

Dark-Glass Viewer
Look at what the machine chose without reading a line of log: previews, edge maps and generated versions of the same frame, cycled with one key.
LVS_Viewer
What is happening

A Qt6 viewer over the ingest and edit databases, with burst grouping, EXIF and sidecar inspection, six score channels, and a variation sandbox that renders soft-edge-guided mockups locally or through an API backend to guide the next pass.

LVS Passive
For people building a dataset rather than finishing a shoot: it keeps a local, searchable record of images and notes made while browsing a web app, so nothing has to be copied out by hand.
LVS_chromepassive
What is happening

A browser extension writes a session record locally; a file-watching daemon copies the referenced images to disk and indexes them in SQLite; a desktop explorer reads it back and exports for training use. It listens only, and everything stays on the machine.

cap height 89baseline 1172113555555
The mark

Drawn to the conventions it is named after.

The field is 233 × 144, a golden rectangle. The three letterforms share one baseline, one cap height and one monoline weight, and the space between them is Fibonacci throughout: 55 wide, 21 apart, 55 wide, 13 apart, 55 wide. The V is the only coloured element in the suite and it is not a letter — it is the two aperture blades of a lens stopped down, which is why it tapers to a single point just under the baseline. The lineage it borrows from is the honest end of optical history: Tessar (1902), four elements; Biogon (1934), symmetrical and low in distortion.

Curved strokes overshoot the cap height and the baseline by a hair so they read level; each glyph is one solid shape. The light gathers inside the letterforms rather than spraying around them, and the V opens or closes to fit the room it is given — which is what it is doing in the header of every page.

The mark, demonstrated

All of it is readable.

Free and open source, no account, no telemetry. Clone it, run it, change it — the licence says you may, and the code is meant to be read.

Browse the repositories