🚀 First INT4 Quantized Efficient-Cube3D - Run on Half the VRAM

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)

B. Medium Categories (7)

C. Complex Categories (5)

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 torchao int4_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)

Shape Quality (Chamfer Distance, 15 categories, 310 prompts):

Median Chamfer Distance: 67.7 × 10⁻³

Best categories: animal_domestic (55.0), vehicle_land (52.2), architecture (54.0).
Complex categories: symmetry_topology (113.6), abstract_mathematical (107.2) — high variance.

Category	Median	Mean	Std	n
Easy (CD × 10⁻³ < 75)
animal_domestic	55.0	60.4	25.5	20
vehicle_land	52.2	61.1	39.0	20
architecture	54.0	61.7	29.9	20
Medium (CD × 10⁻³ 75–100)
musical_instrument	43.1	79.0	86.3	20
animal_wild	65.2	80.4	45.2	20
geometric_primitive	40.8	81.0	90.2	20
furniture	74.9	82.5	39.8	20
fine_detail	57.7	83.3	72.8	20
original_visuals	71.6	79.5	47.3	30
vehicle_air_water	77.8	97.5	78.9	20
Complex (CD × 10⁻³ > 100)
electronics	97.4	126.4	79.3	20
nature_plant	111.3	132.0	69.9	20
tool_hardware	63.7	139.5	193.7	20
abstract_mathematical	107.2	147.4	124.2	20
symmetry_topology	113.6	176.5	169.5	20

Requirements

torch==2.10.0+cu128
torchvision==0.25.0+cu128
torchaudio==2.10.0
torchao==0.10.0

The .pt file is a torchao pickle, torchao enables kernel-supported INT4 inference.

Usage

Please see the

Quantization details

• Method: Round-to-nearest (RTN)
• Precision: W4A16 - weights INT4, activations BF16
• Quantized INT4 layers: 279 / 282
• Skipped layers: shape_proj (in_features=16, < group size), lm_head (out=4099, output head), bbox_proj
• Torchao Quantization Group size: 128

Citation

@article{roblox2025cube,
  title={Cube: A Roblox View of 3D Intelligence},
  author={Roblox},
  journal={arXiv preprint arXiv:2503.15475},
  year={2025}
}