Results: the tree, Blender's tools, the tool choice and simplified pictures

The same tree, made into a body four ways in Blender, on all 37 true models. Pass line, set before the run: the Skin modifier stays unless another tool's mean outline overlap with the true model (6 straight views and a three-quarter view, 1 = identical) is higher by 0.03 or more.

Result: the Skin modifier stays. Mean overlap: the tree's own solid 0.77, Skin modifier 0.75, curves 0.75, metaballs 0.73 (0.66 at a smaller size). Only the solid is ahead, by 0.02, and it is a measuring stick, not an editable model. Metaballs break thin tails and toes into beads; curves leave seams at the forks.

Tree version 2, the same settings for every animal: few points (forks, tips, and where the direction or thickness changes), two thicknesses per point (across and up-down, so flat bodies stay flat), and the body continues only into a clearly thicker branch, so pairs of legs are both branches. Each row: the true model, the tree (balls), the Skin modifier body. Click a row for the side and top views. The grouping below is my reading by eye; turn them in the viewer and judge for yourself.

Bodies with limbs, long bodies and arms: the tree finds every limb, foot and tail, and the Skin modifier makes a clean continuous body.

The parts are there, but a flat box-like slab is left where many branches meet on a thick body (the Skin modifier's known weak spot), or thin legs come out lumpy.

Seven have flat wings, fins or flippers (a tree of round balls cannot hold a thin plate: that is the sheet tool's job); the snail's shell and the jellyfish's bell are hollows; the two anglerfish keep every tooth as a branch; the scorpion is many hard plates; the elephant's thick body leaves slabs.

20 made-up creatures (a body, neck and head, tail, two or three pairs of legs with three to five toes, sometimes open jaws and fins), each with a known tree, scored on 20 more the method had never seen. Pass line, set before the run: at least 95% of the true tips found (one tree tip per true tip), at least 90% of the tree's tips real, and forks within one thickness of where they should be.

Result: not passed. On the unseen creatures the tree finds 91% of the tips (heads, tails and fins all, toes 91%, jaws only 65%: two jaws of an open mouth often merge into one), but 14% of its tips are extra (small side shoots near the feet). Forks land within 0.73 of a thickness: that part passes. (These numbers are after the code review's fixes; the review also found that the scorer skipped any creature whose tree failed, which changed nothing here but did change section 8.)

You asked to see the choice itself tested. 50 made-up parts with a known right tool, 10 each of lump, tube, branched, sheet and hollow (a cup), each in two forms: the true shape, and the shape carved from six pictures, as our pipeline sees it. All five tools were fitted to every part, then three rules picked one:

• Smallest error: the tool whose surface is closest (what we do today).
• Simplest near-tie: among the tools within 10% of the smallest error, the simplest (lump, then tube or sheet, then hollow, then branched).
• Shape reading, as artists do it: read what kind of shape the part is before fitting anything. A cup if much of its hull is empty and enclosed; a sheet if it is thin for its size; a tube if it is long and thin; branched if it has three or more tips; otherwise a lump.

Pass line, set before the run: the right tool on at least 90% of the true shapes and 80% of the carved ones, and at most 10% of last week's passable real parts lost to the choice.

Each strip: the part (gold), then the five tools fitted to it (blue), with their error; the frame marks the one smallest error chose.

The real use: the tree read from pictures, not from a true model. Tried here on the six true models' rendered pictures (six straight views and the four special angles, painted perfectly by part), so the true model can judge. Three ways of reading it, against last week's result (each part's best tool fitted to the carved block). Pass line, set before the run: beat last week's result on the outline from a view no carving used, for at least 5 of 6 animals, and keep 9 in 10 branches.

Result: not passed by any of the three. The best one, a tree per painted part, beats last week only on the coiled snake (0.89 against 0.88 from the held-back view), ties on the scorpion (0.82) and loses on the other four (0.74 to 0.84, against last week's 0.80 to 0.92). By eye it is the most faithful of the three; the numbers favour last week's blocks, which hug the carved outline.

One lesson from the rerun after the code review: on the octopus the tree itself got slightly better (its own solid 0.66 to 0.69), but the Skin modifier broke into a large flat sheet at one crowded junction, and its score fell from 0.83 to 0.74. Skin is fragile where many thick branches meet, which is also where the 37 true models showed their slabs.

The tree from the shape carved from the AI pictures (six straight views and the special angles). The octopus becomes recognisable, which no earlier attempt managed; the agama keeps the slab its disagreeing drawings carve; the jellyfish and scorpion remain tangles.

Artists build insects, crabs and scorpions from separate rigid pieces, one per leg segment and plate, each fitted into its neighbour, not as one skin. I tried the same on our four hard-shelled models. The pieces are the model's own mesh pieces (the scorpion's file is stored as loose triangles, so its corners are joined first). Each piece gets its own small tree and Skin body, and tiny details merge into the piece they touch.

Pass line, set before the run: every piece within one grid cell of the piece it belongs to, so nothing floats.

Result: the outline is closer on all four, and the joins pass on three of four. Crab 23 of 23 pieces joined, spider 21 of 21, scorpion 74 of 74 (two more pieces touch nothing even in the model's own file). On the ant 16 of 19 are joined; three small pieces (1 to 2% of the ant each) sit just over a cell from their neighbour. (These are the numbers after the code review, which found that the first joins test allowed a gap of two cells, and that joining corners at the same place had glued the ant's and crab's separate parts into one.)

Artists block out the large forms first and add small details last, as separate pieces. So I asked the image tool (through Codex) to redraw each reference sheet as a smooth clay blockout: the same pose, cameras and size, with teeth, claws, spines, scales, feathers and eyes removed. This was done for the four true models' rendered sheets and the seven AI animals' sheets, with at most three redraws each.

Pass line, set before the run: every view's outline overlaps the original's by at least 0.90, the details are gone by eye, and the views still agree with each other.

Result: 4 of 11 pass the line, and 10 of 11 look right by eye. The anglerfish, AI agama, octopus and racer snake pass. The others miss mostly because their legs, tentacles or tail are thin, so a shift of a few pixels costs a lot of overlap. The one real failure is the AI seahorse: its body was redrawn wider, as a bottle shape.

Each sheet was carved into a shape (four views), the tree read from it, and both compared with the true model. The anglerfish gains clearly: with the teeth gone, the carved mouth opens, the shape overlaps the true model better in 3D (0.69 to 0.74), and the tree's solid jumps from 0.43 to 0.61 with 19 points instead of 111. The other three lose a little (agama 0.66 to 0.70 in 3D but its tree 0.49 to 0.48; flamingo 0.74 to 0.70; scorpion 0.33 to 0.30): their redraws moved thin legs and plates a few pixels, and the true model they are measured against still has its details.

The tree so far comes from thinning: peel the shape layer by layer down to its centre line. The research turned up two other well-known ways, and I tried both on the same unseen creatures with the same pass line.

• Ball cover (as in Coverage Axis, 2022): the fewest balls that together cover the shape, joined into a tree. It is close to how artists place ZSpheres.
• Mesh shrinking (Au and others, 2008): pull the surface inward, round by round, until it collapses onto its centre line, then read the tree from where it collapsed.

Result: neither beats thinning, and none of the three passes.

The 3D viewer shows ten animals as three models side by side: the true model, its tree (balls coloured by level: body, limbs, toes) and the Skin modifier body made from it. Dragging one turns all three. Fox, chimpanzee, octopus, flamingo, agama, frog, giraffe, deer, elephant and anglerfish: the good, the nearly and the not-yet groups of section 2.