The AI retoucher I wanted would hand me a stack of edits I could still work with. Skin tone on its own layer. Changes to shine on another. Work on the light and shape of a face kept separate from cleaning up a small defect.
I wanted to be able to like one correction, reduce another, and keep the original photograph underneath the whole thing.
That was a specific choice in the original project conversation. When the proposed approach drifted toward passing an image through a generative workflow, I redirected it toward adding layers of edits to the appropriate areas. The interesting output was an editable retouch.
Once you choose that output, the AI problem changes. It has to learn the components of a photographic decision.
The photograph contains several jobs
Evening out skin color and smoothing skin are different jobs. Reducing shine is another. So is altering the pattern of light and shadow that gives a face its sense of depth.
The project separated those operations. The early pipeline included skin smoothing, defect treatment, targeted dodge and burn, skin tone, matte, and volume, alongside other adjustments. That separation is what makes the tool interesting to me as a photographer: the corrections remain things I can think about individually.
Imagine a portrait where the color looks uneven, but the texture and lighting are already right. I would want to work on the color without buying a package deal that also changes the skin's texture and flattens the lighting. On another portrait, the useful edit might be reducing a bright patch of shine while keeping the surrounding light intact.
Those are imagined examples, but they explain the shape of the system I asked for. Each operation should have a reason to exist in that photograph.
The project includes layered PSD output. That connects the automation to the kind of editing document I want: a photograph with decisions attached, rather than a single finished image whose choices are difficult to separate.
Teach from the edits
The other idea came from having retouched work to learn from. In the original conversation, I asked whether those images could become training material. If the examples already contain corrections, perhaps the system can learn something more specific than a general idea of an attractive portrait.
The training work used layered reference files. Those files contain information that a flattened before-and-after can hide: which operation changed which part of the image. A skin-tone layer and a dodge-and-burn layer may contribute to the same finished portrait, but they describe different interventions.
That creates a way to divide the learning problem. Region detection helps identify where an operation belongs. Learned adjustment models predict an operation's correction. The pipeline can then combine those corrections into the image while keeping the operations distinct.
Some parts of the project use learned models and others use conventional image processing. I care more about having the right tool for an edit than giving every step an AI label. Detecting a region, choosing a correction, and composing a layer are useful jobs even when they use different methods.
Predicting an edit instead of a face
The skin-tone work offers a good example of what an adjustment model actually produces. Its target was a correction map, used as a layer over the original. In that representation, neutral gray meant no change. Deviations from neutral carried the adjustment.
So the model wasn't being asked to paint a new face. It was being asked to describe an edit to the existing one.
That is an appealing use of machine learning for me. The photograph remains the photograph; the model proposes a controlled contribution. The same general decomposition makes it possible to work on separate models for different corrections instead of asking one model to express every retouching preference at once.
The creative choice comes before the machinery: what should this operation do, and what should remain available for me to change afterward?
A tool I can keep editing with
My ambition is to make the repetitive parts of retouching easier while preserving the decisions that make a portrait feel right. I want the automation to give me a useful starting point with room to work.
The project had a working layered pipeline and learned-operation work, but the June skin-tone changes still needed retraining and visual evaluation. I wouldn't describe every operation as having reached the same quality level.
What excites me is the direction. Photography supplies the eye for the result. The examples supply specific edits to learn from. AI coding tools and image models make it possible to build a system around those requirements. The result I am working toward is a retoucher whose output still invites me to be the photographer.