Introduction

AssetFurnace 0.4.0 landed on 2026-10-07 at Bingeljell/image-to-3dlab, and it closes the last gap in a fully local character pipeline: type a sentence or drop a picture, and you get back a textured, rigged, animated game character as a GLB for Godot, Unity, Unreal, Blender or three.js. It runs on your own Apple Silicon Mac or NVIDIA PC, with no account, no credits and nothing uploaded.

The project was called Image to 3D Lab until this release. It's built by indie developer Nikhil Shahane (

), the repo code is Apache-2.0, and the repo sits at about 590 GitHub stars as of this writing.

What's new in 0.4.0

  • Auto-rig plus animation, locally. Press Rig and VAST-AI's SkinTokens fits a skeleton and skin weights to your humanoid. Then pick one of 21 preset moves (walk, run, jump, dance, kicks and more), made with NVIDIA's Kimodo motion model and fitted to your character so the rest pose is straightened, arms stay out of the body and feet stay on the floor.
  • The studio. One screen with your Library, a 3D viewer and a Steps panel. + Create takes an idea, a picture or an existing GLB all the way to a moving character, and you can stop after any step.
  • Runs survive a closed tab. Reopen the studio and it picks the run back up; an Activity page keeps logs for every run.
  • A fresh seed on every try, an Open GLB button for models you already have, and a short How to use page.

The full pipeline

Modality: text or image to rigged, animated 3D (GLB). Open vs API: everything runs locally; the app code is Apache-2.0 and each model keeps its own licence. Where to run it today: the one-line installer below, on a Mac with an M-series chip (32 GB recommended) or Linux with an NVIDIA card (24 GB VRAM tested).

StepWhat does itNotes

Text to picture

Qwen-Image 2.1

About 4.5 min on an M5 MacBook, about 20 s on an RTX 4090

Picture to 3D

Pixal3D (recommended), TRELLIS.2, Hunyuan3D-2.1 / MLX ports, Stable Fast 3D

Pixal3D takes about 6 min on an M5 and about 3 min on a 4090

Finish

Retopology, Pixel Match, texture compression

Cuts a typical 30 MB model to under 5 MB

Rig

SkinTokens (MIT)

Humanoids only for now

Animate

21 Kimodo presets

Kimodo model under the NVIDIA Open Model License

Two older features make the output genuinely game-ready. Pixel Match copies your source picture's real pixels back onto every visible surface, so lettering like "VANGUARD 07" on a robot's chest survives even at about 5k faces. Prop sheets (contributed in 0.3.8) turn one picture of a 3×3 grid of props into nine separate, upright, named props, each with three levels of detail.

Install

Mac (Apple Silicon) or Linux:

curl -fsSL https://assetfurnace.com/install.sh | bash

Windows (limited testing so far):

irm https://assetfurnace.com/install.ps1 | iex

The installer is short enough to read first, and it downloads no model weights until you pick a backend in Setup & Status, which shows each size and licence before it fetches. Run the same line again to update. You also need Blender 4.2+ installed yourself for the Finish and rigging steps.

Prompt to try

Rigging works best on a clear, full-body humanoid in a neutral pose, so ask for that up front in + Create:

Full-body stylized game character, a cheerful desert courier with a satchel and wrapped scarf, standing in a relaxed A-pose facing the camera, arms slightly away from the body, simple flat colours, plain white background, no props touching the body

Let it run through picture, Pixal3D, Finish and Rig, then try the run and dance presets and export the animated GLB into your engine.

Caveats

Mind the model licences before you ship anything. The repo's own code is Apache-2.0, but Hunyuan3D weights are not licensed in the EU, UK or South Korea, TRELLIS.2 and Pixal3D bundle Meta's DINOv3 encoder under its own licence, and Kimodo's model uses the NVIDIA Open Model License. AssetFurnace writes a .provenance.json sidecar with the licence of every component for each run, so check it. Rigging is humanoid-only in the studio (a quadruped guide exists for Blender), the TRELLIS.2 and Hunyuan3D-2.1 NVIDIA routes are flagged as not yet tested on real hardware, and Windows support is still early.

Original Source

Conclusion

AssetFurnace now strings together text-to-image, image-to-3D, cleanup, rigging and animation in one free, local studio, which is the whole indie character pipeline minus the subscription. It's still a one-person project leaning on other people's models, so expect rough edges, but if you prototype on a Mac or a 24 GB NVIDIA card, a dancing, engine-ready character from one sentence is worth an afternoon.

—Titus