Cost per Keeper: A Better Way to Compare AI Art Pricing

Every AI art generator advertises the same number: price per image. It is the wrong number to compare. What you pay for is not a generation — it is an image you actually keep, and getting one takes several generations. A tool at half the price that needs three times the attempts is not cheaper; it is 50% more expensive, and it cost you an afternoon. The metric that survives this is cost per keeper.

What cost per keeper means

Cost per keeper is the price of one image you would actually use, after paying for every failed attempt that came before it. Where price per image measures the vendor's unit of billing, cost per keeper measures your unit of value.

Cost per keeper = generation cost × attempts / usable results

Three inputs, each of which you can observe:

• Generation cost — the sticker price. Published, easy, and the only one most comparisons use.
• Attempts — how many generations you spent before something worked. This is where tools differ by an order of magnitude.
• Usable results — how many of those outputs you would actually keep. With batch generation a single run can produce more than one.

Sticker price is one of three variables — and not the one that varies most.

Why price per image misleads

A pricing page can only quote the input you buy. It cannot quote how well the tool listens, and listening is what determines the attempts multiplier.

Consider two tools with a 4× price gap: $0.02 and $0.08 per image. At one attempt each, the cheap one obviously wins. But attempts are not equal. Text-only generators re-roll the entire image every time — change "red coat" to "blue coat" and the face changes too, so a near-miss cannot be nudged, only re-rolled. A reference-based workflow keeps identity fixed and lets you change one variable at a time, which turns a near-miss into a small correction instead of a fresh lottery ticket.

Run the arithmetic at 10 attempts versus 2:

The ranking flipped, and it flipped on a variable that neither pricing page mentions. That gives a rule worth remembering: a generator 4× cheaper per image stops being cheaper the moment it needs 4× more attempts. Substitute your own numbers — the structure holds for any price gap.

Measure your own attempts ratio

Published attempt counts are worthless, including ours: your hit rate depends on what you generate. What is worth doing is measuring your own, once, on a brief you actually care about.

Here is why step two matters more than it looks. The batch below is four outputs from one brief: "standing on a stone ledge, looking over her shoulder at the viewer, full body", in jungle temple ruins.

Score that batch against a strict standard — the pose as written, plus clean hands — and it contains zero keepers: not one of the four is on a ledge looking over her shoulder, and two have visibly wrong hands. Relax the standard to "a good image of the right character in the right outfit and setting" and the same batch contains two. Same four images, same cost, and a keeper rate anywhere from 0/4 to 2/4 depending on a decision I made in my own head.

That is the whole reason to write the standard down first. A keeper rate you define after seeing the outputs is not a measurement, it is a mood — and it will flatter whichever tool you were already rooting for.

What actually moves the attempts number

Four things dominate, and none of them appear on a pricing page:

• Reference vs description. Describing a face in words re-rolls it every time. Anchoring it to a reference image holds identity steady while you change everything else.
• One variable at a time. If changing the outfit also changes the face, you cannot converge — you can only re-roll. This is identity drift, and it is the single biggest attempt multiplier.
• Negative prompts. A standing list of what you never want kills a recurring class of failures permanently instead of per-generation. See negative prompts.
• Batching. Four outputs per run raise the odds that one lands, so attempts-per-keeper falls even though generations-per-run rises.

The fastest way to cut attempts is to stop re-rolling blind: the diagnostic guide to bad generations maps each symptom to the one input worth changing.

Where Gensomnia's numbers sit

Public prices, so you can put them in the formula yourself. Pro is $19.99 per 30 days and includes 500 credits; an image costs 2 credits, which works out to 250 images, or $0.08 per generation. Credit packs run from $9.99 for 150 credits to $99.99 for 2,000 — between $0.13 and $0.10 per image depending on pack size.

That is not the cheapest per-generation price on the market, and the whole point of this article is that it should not be compared as one. The claim worth making is about the other variable: a reference-first workflow exists to cut attempts, not sticker price. Whether it does that for your briefs is a 20-minute measurement, and we would rather you run it than trust us.

Before you pay anything, it is worth knowing what the free tiers in this category actually give you — usually less than the word implies, as this breakdown shows.

What the metric doesn't capture

Cost per keeper is a sharper tool than price per image, not a complete one. Three honest limits:

• It ignores your time. Ten cheap attempts and two expensive ones can cost the same in money and nothing like the same in life. Track minutes alongside the money.
• It is brief-specific. A generator can be efficient at generic portraits and hopeless at a named character in a specific outfit. Measure on what you actually make.
• It says nothing about ceiling. A tool with a great attempts ratio that never produces an image you love loses to an expensive one that does. Cost per keeper compares paths to a standard; it cannot compare standards.

Use it the way you would use any unit-economics number: to kill the comparison that was obviously wrong, not to settle taste. And before you trust anyone else's numbers, run them past the review-auditing checklist — a price without a date and a claim without outputs are not data.