You asked for two things before we try anything else: proper reading on how professional artists build creature models, and a test of how the fitter picks each part's tool, which we had skipped. This page has what the reading says, a first try of the artists' main method on eight of our true models, and the updated plan. The decisions are in chat. The earlier summary is on the roadmap page.
Three new reading sets today, about 110 sources, on top of last week's set on modelling animals part by part (33):
The sources are mostly teaching material and software manuals, so the page states what artists commonly do, not measured results. The papers in the third set have measured results.
The same order shows up in every source, for any animal. The label on each step says where our fitter stands.
Front, side and top pictures, at the same scale, behind the work (Blender: reference images in each straight view; ZBrush: the floor grid).
we do this: the AI picture sheetsDraw the outline over a picture as a curve, or paint it on three walls and keep where all three overlap (ZBrush ShadowBox). That overlap is exactly our carving. Artists use it only as a rough first block, mostly for hard objects.
we do this, but lean on it far moreFor creatures, all sources recommend this. Place one ball per joint over the front and side pictures, each with its own thickness, linked body → limbs → toes (ZBrush ZSpheres, Blender's Skin modifier). The software wraps one continuous skin around the tree. Few balls: a root at the pelvis or chest, the spine first, then about three per limb (shoulder, elbow, wrist) and one per finger or toe joint. Ready-made skeletons for animal groups are often moved and scaled over the pictures instead.
we don't do this yet: the centre of the new planOnly the main masses. Before going on, artists check the outline alone from front, side, top and three-quarter, and shrink the view to see if the shape still reads.
our pictures still carry teeth and claws: the simplify stepEverything that bends shares one skin: body, neck, limbs, tail, wings, fins. Hard things stay separate pieces pushed into the body: eyes, teeth, horns, claws, shells, and every segment of an insect's leg. Separate pieces are fused by voxel remeshing when one surface is needed.
not yet: each part is a separate tool todayTeeth are one tooth copied along a curved jaw; claws, spines and horns are small separate pieces. The mouth is built open from the start, so the inside can be shaped.
not yet: later, after the body works
Made today with the simplest standard method, untuned: thin each true model's solid down to its centre line, read the thickness at each point, drop short branches that barely stick out, merge points that crowd together, and hand the points to Blender's Skin modifier. These are the true 3D models, not pictures: this shows what a tree can hold, not yet what we can get from pictures (that is T3 below). Click a row for the side and top views.
A stricter setting drops more short branches. On the anglerfish it removes some teeth but not the long ones, and on the octopus it drops the whole head, because the head does not stick out far compared with the body's thickness. So details have to go before fitting, in the pictures: your simplify-first idea (T2).
We only ever counted the result. From pictures, the rule picks the tool that is best against the true part in about 4 parts out of 10, and the wrong picks cost 4 of the 18 parts that could have passed. We never looked at how each choice was made, and never tried the rule on parts whose right answer is known. That is the gap you pointed out.
They do not fit everything and compare. They choose by what kind of shape the part is: long and round → a chain in the tree; forked → a fork in the tree; flat → a plate with thickness; compact → a ball or box; cupped → a shell; repeated → one copy, duplicated. With a tree, most of this can be read off the tree itself.
Three habits, all missing from ours. Check against people's answers: part splits are compared with splits drawn by people, and user studies ask whether the result matches what people meant. A shape type that cannot hold the part (a solid tool for a hollow part) is counted as an error, not as a slightly worse fit. Charge for flexibility: a bendier tool must beat a simpler one by more than its extra settings are worth (minimum description length, Leonardis 1997; part-count costs in DBW 2023 and SuperDec 2025). Report the gap to hindsight: Kobsik and colleagues (2026) choose among 45 candidates by lowest error and reach 65.6% overlap, where the best in hindsight would give 74.4%. We have the same gap.
Two changes. The choice test now comes first, alongside the clean truth tree and the simplified pictures. And the fitter now copies the artists' main method: draw a tree of balls over the pictures, then let the Skin modifier make one body, instead of carving a block and fitting a tool per part. The five tools stay, but only for what a tree cannot hold: fins and wings (sheet), shells and bells (hollow), and separate hard pieces.
Done today: the reading, and the first tree on eight true models (above). Still to do: the five briefs, the made-up shapes for T0 and T1, and the lab data staged.
T0, T1 and T2 start together: three Codex workers, and image calls for T2. T5 starts when a worker is free.
T3, then T4, built on the clean trees, the simplified pictures and the chosen rule. Each result comes to you as pictures.
The decisions are in chat.