Deemos Teaches Hyper3D Rodin to Build Animation-Ready Low-Poly Meshes in 20 Seconds
Meshes from neural networks usually have to be rebuilt by hand: a chaotic mesh bends poorly in animation. Deemos has updated the Low-poly mode in Hyper3D Rodin. It builds a lightweight mesh of 200–10,000 faces with rings of edges around elbows and knees. A textured mesh is ready in 20 seconds via the API or an MCP server for AI agents. For indie studios, that means hours less manual retopology.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Neural networks have long been able to produce good-looking 3D models, but you usually can't drop such a model into a game or animate a character with it: inside, it has a chaotic mesh of many tiny polygons (faces). Picture a hero whose elbow crumples into an ugly accordion when it bends, while the model itself is so heavy that it slows down a mobile game. To fix this, an artist does retopology by hand, rebuilding the mesh so that it is clean and economical.
Deemos Tech has updated the Low-poly mode in its Hyper3D Rodin model so that an algorithm takes over this routine. Low-poly is a lightweight mesh whose face count the user sets, from 200 to 10,000. "Clean topology" means the edges run correctly, in rings around the joints.
The whole process, textures included, takes about 20 seconds, and you can launch it through an API (programmatic access from your own code) or through an MCP server (a standard way to connect a service to an AI agent).
What the new mode can do
The main idea is a mesh you wouldn't be embarrassed to hand to an animator. The model builds clean edge loops around joints and bend points. This is the layout without which a knee or elbow deforms badly during rigging, that is, when the "skeleton" is set up for animation. If the loops run around the joint, the bend looks natural.
The second parameter is the polygon budget, meaning how many faces you are willing to spend on an object. It can be set from 200 to 10,000, and the mesh can consist of quads (four-sided polygons), triangles or n-gons (polygons with any number of corners). You won't have to convert anything by hand for different engines and pipelines. A couple of hundred faces is enough for a mobile game, while for a hero close-up you can go up to the ceiling of ten thousand.
So that the lightweight mesh doesn't look like plasticine, details from the original sculpt are transferred to it through normal baking. Put simply, small wrinkles, folds and seams are not in the geometry but "painted" into a texture that is lit as if the relief were real. This is a standard technique in game graphics, and the author of the @cgevent channel notes that everything here is done "the way the grown-ups do it."
How to run it
Generating the mesh and textures takes about 20 seconds. There are two ways to access it. Through the API, you can build the feature into your own service. Through MCP, Rodin connects to an AI agent or an assistant in an editor: the agent itself requests a low-poly version of an asset at the size it needs, and a person doesn't have to open the web interface.
In the PolyGen API, the target face count is also set from 200 to 10,000. The input can be:
- a ready-made Rodin task already created in the service;
- an uploaded model of your own;
- from one to five images;
- a plain text prompt.
The supported formats are OBJ, GLB, USDZ, FBX and STL. If you upload your own model, textures are carried over from the original, so colors and materials don't get lost during simplification.
What this means in practice
For indie studios and small teams, this cuts down on routine work. The chain of "generate a model, hand it to an artist for retopology, bake normals" used to take hours. Now a rough version good enough for testing in an engine takes about as long as pouring a cup of coffee. For prototypes, AR scenes on USDZ and mass-populating levels with props (furniture and set dressing), this approach looks especially sensible.
Product teams should take a look at the API and MCP combination. If you are building a game-creation tool, a product configurator or an avatar generator, low-poly export can be built in as a ready-made step rather than a separate stage for people.
There are caveats. For now, this is a claim by the developers and a retelling by the author of a Telegram channel; the source contains no independent measurements of topology quality. The stated 20 seconds probably depend on model complexity and service load. A ceiling of 10,000 faces suits props and game characters but is not enough for cinematic quality. Loops around joints work best on humanoids and familiar shapes, and on unusual creatures the result should be checked by hand.
A sensible plan: take a couple of your typical assets, run them through PolyGen at different budgets and see how the mesh behaves in a rig. If even half of the models pass without fixes, retopology will stop being a bottleneck in your pipeline.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
What the new mode can do?
The main idea is a mesh you wouldn't be embarrassed to hand to an animator. The model builds clean edge loops around joints and bend points. This is the layout without which a knee or elbow deforms badly during rigging, that is, when the "skeleton" is set up for animation. If the loops run around the joint, the bend looks natural. The second parameter is the polygon budget, meaning how many faces you are willing to spend on an object.
How to run it?
Generating the mesh and textures takes about 20 seconds. There are two ways to access it. Through the API, you can build the feature into your own service. Through MCP, Rodin connects to an AI agent or an assistant in an editor: the agent itself requests a low-poly version of an asset at the size it needs, and a person doesn't have to open the web interface. In the PolyGen API, the target face count is also set from 200 to 10,000.
What this means in practice?
For indie studios and small teams, this cuts down on routine work. The chain of "generate a model, hand it to an artist for retopology, bake normals" used to take hours. Now a rough version good enough for testing in an engine takes about as long as pouring a cup of coffee. For prototypes, AR scenes on USDZ and mass-populating levels with props (furniture and set dressing), this approach looks especially sensible. Product teams should take a look at the API and MCP combination.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.