Presenting the first INT4 quantized efficient version of Cube3D v0.5, a text-to-3D mesh generative model.
Quantized via RTN W4A16 (group_size=128) using torchao, it cuts the model size from 7.2GB → 1.3GB (82%↓)
and peak VRAM from 25.4 GB → 11.3 GB (55%↓) while maintaining the same inference speed and comparable shape fidelity -
enabling 3D shape generation on much smaller, more accessible GPUs.
BF16 + Engine
BF16 + EngineFast
INT4 + EngineFast
💾 Model size
7.17 GB
7.17 GB
1.26 GB (82%↓)
🎮 Peak VRAM
21.7 GB
25.4 GB
11.3 GB (55%↓) ✨
📦 Setup time
19.4 s
206.9 s
6.9 s (97%↓)
⏱️ Latency
90.9 s
15.0 s
14.2 s
💡 The 82% size reduction and 55% VRAM reduction means this model now fits on a single 15 GB GPU (e.g. NVIDIA L4, A10, A2 etc.), bringing high-quality text-to-3D generation to individual researchers and end-user hardware.
Original BF16 vs Quantized INT4 Comparisons:
A. Easy Categories (3)
Easy categories
B. Medium Categories (7)
Medium categories
C. Complex Categories (5)
Complex categories
Cube3D v0.5 - RTN W4A16 INT4 (torchao)
Post-training quantized version of Roblox/cube3d-v0.5, a text-to-3D mesh generative model.
Quantization method: RTN W4A16, group_size=128, via torchaoint4_weight_only.
What's in this repo
File
Size
Description
shape_gpt_rtn_int4_g128.pt
1.26 GB
INT4 quantized GPT weights (torchao pickle)
shape_tokenizer.safetensors
~1.10 GB
VQ-VAE decoder — BF16, unchanged from base model
open_model_v0.5.yaml
tiny
Model architecture config
quant_config.json
tiny
Quantization metadata
New Benchmarking Dataset (15 categories, 310 prompts)
-blueviolet){kind=icon}
-brightgreen){kind=icon}
-blue){kind=icon}
-yellowgreen){kind=icon}
