-
Notifications
You must be signed in to change notification settings - Fork 14
Help Train From Edits
The Save Edits as AI Training Examples workflow lets you teach LrGeniusAI your personal editing style. Your current Lightroom develop settings for selected photos are saved to the backend as labeled examples. They are what AI Edit Photos builds its recipes from — without them AI Edit has nothing to work with and refuses to run, so this workflow is the prerequisite for that one, not an optional refinement of it.
Library → Plug-in Extras → Save Edits as AI Training Examples...
- Edit a set of photos in Lightroom's Develop module the way you like them.
- Select those photos in the Library grid.
- Open Save Edits as AI Training Examples from the menu.
- In the dialog, choose the scope and optionally set a style label (e.g. "Wedding summer 2025") and a description of the style. Both are optional.
- Confirm.
The plugin reads the current develop settings from Lightroom and sends them to the backend, where they are stored as labeled training examples.
- Selected photos only — reads develop settings from your current selection.
- Current view — uses all photos in the currently visible folder or collection.
- Entire catalog — reads develop settings from all photos in the catalog.
A short name for this set of training examples (e.g. wedding_airy, street_bw, landscape_dramatic). Use the same label when saving multiple sessions of a consistent style.
A free-text summary of what this edit style represents. Helps the AI understand when to apply it.
Training examples are the entire input to AI Edit Photos. The backend finds the closest examples by image similarity, re-scores them on exposure, scene type and time of day, and interpolates their develop settings into a recipe for the photo in front of it. No language model is involved.
That has two consequences worth knowing:
- Below five examples, AI Edit will not run. It stops in the plugin with a pointer back to this workflow. Five is the floor; the style profile is reported as Building up below ten and Active from fifty, and the match quality follows that curve.
- Variety matters as much as volume. Matching is per photo, so a profile built only from bright studio work has little to offer a night shot. Train on the kinds of photos you actually want edited.
Matching runs on image similarity, so a training photo has to have been indexed with Enable smart photo search for it to be findable. Saving an example for a photo without an embedding still works, but that example will never be retrieved — the plugin shows a warning when this happens.
The same applies to the photo being edited: if it is not indexed, there is no embedding to compare your saved edits against. The backend then falls back to scoring recent examples on exposure, scene and time of day alone, which is a weaker match — the review dialog reports the lower confidence rather than hiding it.
Only for white balance, and only because Lightroom forces the issue: its temperature slider is Kelvin on a raw file and a relative −100..+100 on a JPEG. Averaging an example edited from a raw with one edited from a JPEG would produce a number that means nothing on either scale — and it would land on a real photo.
So when the backend blends a white balance, it uses only the examples whose file type matches the photo you are editing. If none of the matched examples qualify, it leaves white balance alone and tells you why. Everything else — exposure, contrast, presence, colour, curves — means the same on both, and is blended from all matched examples as usual.
Practically: if you shoot raw and JPEG both, keep saving examples from both. Nothing is wasted; the pool simply narrows for that one field. Examples you saved before this existed count as compatible with anything, so nothing you already have stops working.
- Save at least 5 well-edited photos per style — that is the Style Engine's minimum — and 10–20 for good results. More diverse examples generalize better.
- Use consistent labels per shooting genre: one label for weddings, one for portraits, one for landscapes — rather than mixing everything.
- Re-run the workflow after editing sessions to keep examples fresh and representative.
- Training examples are not one input among several — they are the only one. AI Edit has no look presets, no style-strength slider and no prompt to fall back on, so whatever the examples say is what the recipe does.