PresetForge vs asking a chatbot to write the XMP
Ask a chatbot for a "moody teal-and-orange preset" and you will get a well-formed XMP file in seconds. Ask it to reproduce the edit in these two photos and you will get a well-formed XMP file in seconds. Only one of those answers is meaningful.
Where asking a chatbot for xmp values wins
- Excellent at explaining what a slider does and why a look works.
- Fast at producing a plausible starting point from a description in words.
- Free or near-free, and already open in another tab.
- Good at the boring parts: XML structure, naming conventions, batch-renaming a folder of presets.
- Genuinely useful as a tutor — it will explain the curve better than most tutorials.
Where measuring wins
- Reads the actual pixels. The numbers come from the image, not from a plausible guess about images.
- Per-channel curves fitted to real correspondence, with a reported residual error you can check.
- Deterministic: the same pair produces the same preset every time.
- The output is verified against the source — you can see how closely the derived preset reproduces the target.
Use a chatbot to understand the tools and to draft from a description. Use measurement when there is a specific edit that has to come out right.
What a language model can and cannot see
A model that can look at images can describe one accurately: warm, contrasty, lifted blacks, greens pulled toward yellow. That description is real information and often a good one.
What it cannot do is measure. Producing an accurate tone curve requires comparing tens of millions of pixel pairs and fitting a monotone function through the medians of each brightness bin. That is arithmetic over the full image, not something that can be estimated from a description, however good the description is. Ask for specific numbers and you will get numbers that are typical of the kind of look you described — which is a different thing from the numbers that describe your look.
The confidence problem
The failure mode that costs people time is not that the answer is wrong. It is that the answer is wrong and well-formatted. A generated .xmp will have the right namespace, plausible crs: attributes and a sensible name, and it will import cleanly. Nothing about it announces that crs:Contrast2012 was chosen because contrasty photos usually have contrast around there.
A measured preset can be checked. Apply the derived curves to the original and compare the result to the target: the mean absolute residual per channel tells you, in units you can reason about, how close the reproduction is. PresetForge reports that number on every extraction. There is no equivalent for a guessed preset, because there is nothing to compare it against.
Where a chatbot is the better tool
Understanding. If you want to know why lifting the blue channel in the shadows produces the look it does, or what the difference between the parametric and point curves is, asking is faster than reading and you can keep asking follow-up questions. That is a real advantage and no amount of measurement replaces it.
Drafting from a description, too. "Give me a starting point for a washed-out summer look" is a request that has no measurable answer — there is no source edit — so a plausible guess is exactly the right kind of answer. Take it as a first draft and tune it.
And the mechanical work around presets: writing a script to rename a folder, explaining why an import failed, converting a settings list between formats. Models are good at that.
A note on cost and determinism
A measurement runs the same arithmetic every time, so the same pair of images always produces the same preset. That matters if you are building a library and want consistency across a hundred extractions. It also means there is no per-request inference cost, which is why the free tier can be five real extractions a month rather than a trial.
If you want to understand what the numbers mean before you trust any source for them, what tone curves actually do is the shortest route to being able to look at a preset and tell whether it makes sense.