How Helen Gets Made A making-of, not the live show Watch her live

Workshop · the AI models · 3D and image generation · the bill

What it cost and what it saved

Three hundred and forty credits, eight tasks, five files kept. The cat exists because of the generator and looks like her because of everything that happened after.

The tasks are itemised on the Tripo page; the hours went into the shippable-mesh page.

Helen, a dilute tortoiseshell calico, the real cat every generated model and picture on these sites is measured against
A week of Blender for someone who knows Blender, or seventy credits and three minutes plus the work on the other four pages.

Tripo bills in credits and the plan behind this account is not written down anywhere on this site, so there is no dollar figure here. There is a credit ledger, a clock, and a comparison with what the alternative would have been for one person with a day job and no Blender. The tasks themselves are itemised on the Tripo page.

The ledger
Spent onCreditsOutcome
The court's cat, September (views + mesh)10 + 60on the ottoman now
The court's cat, August, standard quality30served for sixteen days; still linked as a download
The same photo at detailed quality606.3 MB file, wired for wide screens, never served
The standing pose from two real photos30the fallback if the sitting file fails
A first sitting cat from an album frame30replaced within hours
The two Morse busts60 + 60both on the about page
Total, eight tasks340five files live, three discarded
The clock
JobTimeWhose
View task posted to a packed, committed, shippable September cat16 minutestwo Tripo tasks about three minutes of it; the rest the pack script, a look, and a commit
Each August generation, upload to download2 to 5 minutesTripo
Writing the Tripo scripts, the two pack scripts and the fit codenot clocked; the notes are dated 22 and 23 Augusta coding assistant in the repository, a person reading and deciding
The tongue repair30 minutesa person, with two twenty-line throwaway tools
The KTX2 experiment, run and rejectednot clocked; one header comment survivesa person looking at fur
Judging each candidate against the twelve checksminutes per candidate, repeatedlya person, every time

Where generation won

The cat exists. The first argument stops there, and it is a strong one. A walk-around, textured, sit-posed likeness of a specific cat is a week's work for someone who knows Blender and can sculpt fur, and the operator is not that person and was not going to become one for a cat. Seventy credits and three minutes produced a mesh that passes the only check that matters to her owner, after the fit and the pack. The fallback had been a procedural figurine built from tapered tubes and a painted patch map; it looked like a nice toy.

The busts are the second win. Two historic photographs of two real people became sculptures in an afternoon for 120 credits, and the flaw in one of them, a whole figure where a bust was asked for, was fixed with one clipping plane and one number. Commissioning two busts would have been a different order of cost and would have taken longer than the entire site did.

The four synthetic views are the third and the cheapest. Ten credits bought a back for the cat. There is no manual equivalent short of photographing her from behind, which anyone who has tried to photograph a cat from a chosen angle will understand is not a manual equivalent.

Where it did not

Three of the eight tasks were discarded, 120 of the 340 credits. One was an album frame that produced a cat nobody kept. One was a higher-quality version of a file that already looked right, generated on the theory that wide screens deserved more and never served because the standard file was enough. Generation is cheap enough that trying is the default, and cheap enough that the discards do not sting; they are recorded here because a page that only lists the successes is advertising.

On the 2D side the record is worse. The channel kit's generated cat cost a day of design work to replace, and the court's rear-wall portraits are being replaced now. Both were rejected for the same reason: they were competent pictures of some other cat. The generated still that has stayed, the Royal Portrait, stayed because a person chose it from twenty for looking like her.

What a credit buys, in this project's units

  • Ten credits: four views of a cat from one, three of them invented.
  • Thirty credits: a standard-quality mesh, 48,000 triangles, from one image. Fine at court distance.
  • Sixty credits: a detailed mesh from one image or four views, up to two million triangles. Needs packing before it is usable.
  • Zero credits: every re-run of the pack script, the fit code, and the look. Those are the expensive part, measured in hours rather than credits, and they are the part that decides whether the result is her.

If you are costing your own

Budget the credits for at least three attempts per subject, because the first photograph is rarely the one that works and a regeneration is a new mesh, not a retry. Budget a person's afternoon for the fit code the first time and an hour per model after. Budget for one repair per model that no script can do. And decide before you start what the acceptance test is; here it was a single question, "is it Helen?", and a person asked it of every file.

The work those hours went into is on the shippable-mesh page; the pictures that were paid for in design time rather than credits are on the site-imagery page; and how the other models on this build earn their keep is on the AI page.