225GB of Free 3D Models, Compressed to 3.2GB, and Doors Still Open in the Browser

An Xuan converted NVIDIA's 2,595 free SimReady models on Hugging Face into GLBs that browsers can load directly. The original assets exported from Blender were about 225GB; after conversion, the whole package is 3.2GB to 3.8GB. For the V8 engine, a single file dropped from 888MB to 7.4MB to 9.3MB, and Chrome's first frame went from more than 9 seconds to about 0.6 seconds. He recently used Opus 5.5 to build an interactive warehouse in the browser using this asset set.

There are two versions of the numbers. The tweet says 225GB to 3.8GB, while the dataset page was later updated to 228.1GB to 3.2GB. In the tweet, the V8 engine is 9.3MB and 0.6 seconds; the dataset page says 7.4MB and 0.57 seconds. The counting time and compression level differ, but the order of magnitude is consistent.

The common problem with free 3D assets is that you get a download link and your hard drive immediately complains. 2,595 models, 821M triangles, max texture edge 4096 pixels. Loading them directly into the browser is unrealistic. An Xuan's pipeline first downloads SimReady USD, imports it into Blender 5.2, removes collision, proxy, and grasp-assist geometry, and exports GLB. Then it uses glTF Transform 4.5 for deduplication, cleanup, and welding, compresses textures to 1024-pixel WebP, applies high-level meshopt compression to geometry, uses octahedral encoding for normals, and quantizes attributes.

The key point: no mesh simplification was done. Triangles dropped from 821M to 415M by removing duplicate instances and helper geometry. The models' appearance did not change. The dataset page provides 2,595 thumbnails for side-by-side comparison. The conversion scripts are open-sourced on GitHub.

Physics data is the focus of this update. 2,587 files have rigid bodies, mass, and physical materials; 929 files have joints, 5,550 in total. The tweet says 5,546, a difference of 4, possibly due to different counting times. Door hinges, drawer slides, buttons, wheels, and lids are all there. The data is stored in two copies: one in Khronos' KHR_physics_rigid_bodies extension, and one in node extras, which three.js puts into object.userData. In the viewer, every movable part has a slider; doors can rotate and drawers can be pulled.

The limitations are also clearly stated. SimReady originally used signed distance fields for collision; in GLB they are replaced with convex hulls, so hollow cabinets and baskets collide like solid objects. KHR_physics_rigid_bodies is still a Khronos review draft, so you need to check for yourself whether an engine supports it. Joint drives and USD simulation settings were not carried over, and MDL material effects are lost. This file set is better suited for display, lightweight interaction, and prototyping, not as a direct replacement for high-fidelity simulation.

To use it, the dataset is at codepawl/simready-glb, the whole package is about 3.2GB, and you can download by category. To load it in the browser with three.js, you need to set MeshoptDecoder, because the geometry uses EXT_meshopt_compression and the textures are WebP. The online 3D viewer can search all 2,595 models and rotate them directly for viewing. The conversion scripts are on GitHub. The model author is NVIDIA SimReady, licensed under CC BY 4.0; please credit the source when using it.

What is truly scarce about free 3D assets is being able to go straight into a workflow. The quantity is already sufficient. NVIDIA provided the models; An Xuan filled in the size, format, and physics data. What remains is to see who uses it to make something. The original tweet is here.