ByteDance-Seed/SimArt

Brief overview

ByteDance Seed/SimArt model pre-training weights released, including multi-modal large language model (MLLM) weights for generating joint-based 3D object structures (URDF files), and sparse VQ-VAE model weights for encoding and decoding 3D part geometries into discrete latent spaces.

Article

SIMART Checkpoints

This directory contains the pretrained weights for the SIMART models.

Directory Structure

`simart_mllm/`

Description: This directory contains the weights for the Multi-Modal Large Language Model (MLLM).
Base Model: Fine-tuned from Qwen3-VL-8B (or similar Qwen-VL architecture).
Purpose: Used for generating articulated 3D object structures (URDF files) based on multi-modal inputs.

`simart_vqvae/`

Description: This directory contains the weights for the Sparse VQ-VAE model.
Purpose: Used for encoding and decoding 3D part geometries into/from discrete latent spaces.
Contents:
- encoder.pt: Weights for the VQ-VAE encoder.
- vq.pt: Weights for the Vector Quantizer (codebook).
- decoder.pt: Weights for the VQ-VAE decoder.