PerfuMate › AI Studio
It reads your shelf, not the internet
Describe what you want and AI Studio returns a complete formula, in grams, every line from a bottle you already own. What separates it from asking a chatbot for a perfume recipe is not the writing. It is that it is handed our catalog on every request - the odour description, the strength, the skin life and the dose ceiling for each of your materials - and that nothing reaches you until the arithmetic has been checked.
Five things it does
- BuildsA brief in, a complete formula out: pillars, the middle band, and the trace tail that makes it that perfume and not a generic one. Perfume or accord. It arrives named, with a percentage and a reason on every single line, and it will tell you plainly when what you asked for would make a weak perfume.
- AdvisesOpen a formula you already have and ask what is wrong with it. You get a real critique with numbers, not encouragement.
- RevisesA second tap turns that advice into a V2, saved as a new copy, with the original untouched. Advice you have to retype by hand is advice you will not use.
- RebuildsHand it a community formula, or a PDF from anywhere, and it recomposes the same idea from the bottles you actually own, saying for each swap how close it gets.
- AnswersAsk it a craft question and it answers as a perfumer: concrete percentages and material names, or one line if one line is honest.
What it is handed, bottle by bottle
The model is not trained on our catalog. It is given it. That distinction is worth being precise about, because the second thing is better than the first. A trained model carries a blurred memory of what it learned months ago. This one is handed the live rows, per request, for the materials on your shelf - so when the catalog is corrected, the next run is corrected with it, and there is no version of the data drifting along inside the model where nobody can reach it.
Every material you own travels with the following, up to a 350-material working set:
The written odour description, in fullodor | Not the first line of it. The second half is where the bench detail lives: what a material fights with, what it does at trace against at dose. That is what makes an answer fit a brief instead of merely sounding plausible. |
Its olfactive family, and its facets under thatfamily | Placement down to facet level, so "smoky" can mean burnt cedar or pipe tobacco and the difference survives the trip. |
Odour strength, 1 to 10strength | Eight and up are trace-dose materials. Ten is the class where a single drop takes a formula over. |
Its own hard ceiling, as a number on the rowmaxPctOfConcentrate | Potent materials carry the exact figure the server will judge the answer against, in the same unit the answer comes back in. The model is not asked to derive a limit from a rating. |
How long it lasts on skin, in hourssubstantivityH | This is the longevity arithmetic. A twelve-hour brief is engineered from these numbers, not guessed from the word "base". |
Top, heart or basevolatility | The tier it naturally occupies. |
Its registry numbercas | How the IFRA limits are matched. A registry number has no spellings, and a name does: this is the join that stops a real limit being lost to "Phenyl Ethyl Alcohol" against "Phenylethyl alcohol". |
The dilution you keep it atworkingDilutionPct | And every bottle strength you own of it, so the advice matches your weighing practice. |
Where it sits on your shelf for costpriceBand | Cheap, mid or expensive, as terciles within your own shelf rather than a market price we do not know. |
Whether that bottle is nearly emptylowStock | Fine for a trace; for a structural pillar it says so, or picks something else. |
Behind those rows sits the catalog itself: 1,581 materials, every one of them carrying a written odour description, 1,153 with a registry number, 196 flagged as restricted. Every odorous material carries a skin life, 441 of them from a sourced figure rather than an estimate from its family. You can read the whole thing without an account, and it is the same data the model gets.
Scent direction: the studio's equalizer
The brief is prose, but the direction is dials, and they steer every message you send rather than just the next one. Nothing here is required - leave it all on Auto and you get a perfumer's own judgement.
Families, and their sub-notes
Tap a family to steer toward it; hold it and you pick the facets underneath, so "woody" can be sandalwood-creamy or cedar-dry.
Longevity target
Intensity
Creativity
And two switches
"Only materials I own" is on by default, and it is not a hint: it is enforced on the way back. Turn it off and the wider catalog opens, and every off-shelf line has to argue for its own purchase.
The Scent Equalizer
The other way in. Open a formula, say where to take it - more mossy, less smoky, down to the facet - and the answer is not a guess. It is arithmetic on your own shelf. The engine knows what fraction of each of your materials actually lands on the note you moved, and solves for the grams that put the share where you asked:
x = M(t − s) / (w − t)
A material that only partly carries that note needs more grams to move the same distance, and one that cannot reach the target however much you add is told to you as such, with where it does land. Then the studio retunes the formula around the move - and the brief it writes lands in the box first, in plain English, editable, before anything is spent. You read the sentence you are paying for.
The drydown is arithmetic, not a mood
Ask for twelve hours and something has to actually answer for it. Every material's skin life rides in the payload in hours, and it is the skin figure - paper holds a scent far longer than skin does, and the two are not interchangeable. Attach a formula and the app's own reading goes with it: the family shares, the split across the pyramid, the longevity estimate and the IFRA verdict, all computed before the model is asked anything. It reads what the bench already worked out instead of estimating it a second time, which is also why its answer and your drydown chart agree.
Nothing reaches you unchecked
A language model under a creative brief will quietly break a rule stated in prose. So the rules are not prose on the way back. Every formula is validated before it is shown to you, and a failure sends it back to be done again inside the same credit:
- The shelf is enforced, not requestedWith "only materials I own" on, a line naming anything else is rejected outright. The model cannot invent a bottle you do not have.
- Potent materials are held to arithmeticA strength-ten material has a hard ceiling in percent, and the line is judged neat-equivalent - weighed from its own bottle strength, and for an accord, scaled by the dose that accord is used at. This exists because the model once overdosed a material a perfumer caught by hand.
- IFRA limits are walked all the way to skin291 verified Category 4 rows, matched by registry number rather than by name. A line is taken out through its bottle strength, the accord's share and how strong the finished perfume is worn, and then compared. Isomers that share one limit are totalled as a group rather than each getting the full allowance.
- A restriction we cannot quote is treated as a small oneWhere a material is flagged restricted but has no verified number to send, it is held to a trace rather than given the benefit of the doubt. A limit nobody can quote is a reason to dose small.
It is a reference, not a certification. The IFRA layer is built against the 51st Amendment for fine fragrance. It is there to catch the mistake you did not know you were making, and it is not a substitute for the standard itself or your supplier's documentation. The same is true of everything else here: this is a bench tool, and you are still the perfumer.
What it will not do
It does not browse the web. Ask for something in the style of a commercial perfume and you get an interpretation built from your shelf, presented as an interpretation, never as the original formula, which we do not have and would not print. It will not state an IFRA number from memory: if the verified figure is not there, it says the restriction and gives no number. And it does not flatter. If the brief would make a weak perfume, it says so and proposes the better version alongside what you asked for.
What it costs
The bench is free, and it is complete without any of this. Every calculation, the whole 1,581-material catalog, the drydown, the equalizer's own arithmetic, the evolution tree, export - none of it needs a run. AI Studio is priced on its own because every run is a live bill to the model on the day it happens, which makes it the only part of PerfuMate that renews. 5 runs come with the one-time purchase; after that:
Yearly is cheaper, unused runs roll over up to two months of your plan, and what you hold when a plan ends stays yours. The full price list has the rest.
More on the same shelf: the catalog the model is handed, what happens to your data, what everything costs, and the beginner's guide if you have not made a perfume before.