One image, 260 parts: He packed 3D disassembly into a Claude skill
Elvar used a single reference image to have Opus 5.5 generate a complete 3D disassembly: three rounds of comparison renders, 260 parts hidden beneath the gears, and one part removed every 2–3 seconds. He made seven in two days, and both the skill and 9 prompts are public.
Elvar packed an entire 3D workflow into a Claude skill. Here, a skill can be understood as a reusable workflow package that locks down prompts, steps, and constraints so Claude can use them repeatedly.
The process is short: give Opus 5.5 a reference image and ask it to generate a complete disassembly based on the image, with every part written in code. The appearance is determined by the image; the shell, gears, and hoses are all three.js code.
The interface in the video is the Mellis HB-7 bionic flyer: six legs plus one wing, with a control panel running data such as wingbeat frequency, power, battery level, and flight mode. It looks like real-time monitoring, but behind it is a model assembled from code.
Key details are below. Before producing the video, Opus did three rounds of side-by-side rendering comparisons, putting the render results next to the reference image for comparison. Hidden beneath the large gears are 260 tiny pistons, hoses, screws, and wires. The first version could only hover; Elvar changed it so that a part separates every 2–3 seconds; pull one leg, and its servo rod flips to the DETACHED state.
In two days he made seven such sims and put them in a playground. The skill, the playground, and the 9 prompts are all public, with links at the end.
The comments show two kinds of reactions. One camp thinks, “Going from a single reference image all the way to code for the whole thing—this workflow is pretty hardcore.” The other is more technical: some users ask whether 260 moving parts can hold a stable 60fps, or whether the three rounds of matching only fixed the appearance while performance has to be handled on your own. This question is worth looking at more than the “wow” factor. A generative model writing out appearance and structure is one layer; real-time interaction and performance are another. Elvar’s demo pushes the first layer to a very concrete point: it can pull a leg, flip states, and separate one part every few seconds. Things that stop at the concept-image stage cannot do this.
The source material doesn’t give the number of hours for each build, nor does it say whether 60fps was achieved. What can be confirmed is that the 7 sims are live, and the skill and 9 prompts are available. Anyone who wants to reproduce it can first look at the part density in the video, then see whether their own machine can handle it.
Link: https://x.com/elvar_official/status/2116192984299344372
发布时间: 2026-10-31 10:53