Four steps.
No pipeline diagram.
This page follows one frame from the card to the finished file, the way it happens on a Tuesday night. Numbers where the numbers are interesting. Plain words everywhere else.
The screen fills before the drive does.
The previews baked into your files are read first and the full RAWs are copied underneath them, so the frames are in front of you in the time it takes to make tea while the transfer finishes on its own. Nothing about this step waits for the interesting part to be over.
Same viewer, same keys, one extra line on screen.
Whatever program you trust is still the program you are in. An overlay shows what was noticed about the frame on screen — 04/10, backlit, dress reading as blown — and 1–5 files it and moves on. The ratings land in your EXIF, so Lightroom and your archive see exactly what they saw before.
You describe the edit; a recipe comes back.
Talk to it in the chat window you are already in, or hand it the whole folder and walk away. It is not painting your photograph — the model writes settings, your machine renders them, and the settings are a normal RawTherapee file at the other end. It works the same on a building: “keep the stone cool, the windows can blow.”
Skin, eyes and backdrop get their own masks.
Local work is where most edits fall apart: a warm face against a cold window, a highlight running down one cheek, a sky already at its limit. Those areas are separated and adjusted on their own, so the sky keeps its gradient while the face comes up — and the result leaves as an image with the layers still in it.
The batch runs while you are not there.
Give it a folder on your way out. Strong frames come back finished, doubtful ones come back flagged with the reason — “face in shadow, no shadow detail recovered” — and none of them are quietly guessed at. In the morning you approve, or you say why not.
Where the machine stops and you start.
Reading a frame: what is skin, what is fabric, what is already clipped, which way the light came from.
Turning an instruction into values, in the right order, for this camera and this file.
Finding the edges of a subject so a local adjustment lands on it and nowhere else.
Working through hundreds of frames overnight without asking you to sit and watch.
What the photograph is about — the frame worth keeping out of four that look almost identical.
What the light means. Warm, honest, cool, moody: not a slider position, a decision.
How far is too far on skin, and when a subject should stay as found.
The last look. Approve, reject, or say why — that note is the next instruction.
It will, and it is built on that assumption.
A frontier vision model is confidently wrong about something roughly as often as it is right about something difficult. So the suite is arranged around a photographer checking its work rather than around trusting it. Nothing is overwritten in place, every recipe is reversible, and every decision it made is visible next to the frame it made it for.
What it cannot do is argue with you. The star you gave a frame is the record; the next pass starts from your taste, not from its own.
A preset is one answer, applied to every photograph that will ever be made in that style. It cannot know that this frame was shot into the sun and the last one was not.
LVS looks at the frame it has, in the light it was actually shot in, and decides what that one photograph needs. Ask it the same thing tomorrow about a different frame and you will get a different answer — because the frame is different.
Presets assume every face is lit the same. Nothing is.
Where the pieces live.
Seven repositories, one job each, all readable, all replaceable.