ilessio-aiflowlab/project_huginn

HUGINN — OMG4 Optimized Minimal 4D Gaussian Splatting

Part of the ANIMA Perception Suite by Robot Flow Labs.

Paper

OMG4: Optimized Minimal 4D Gaussian Splatting (Oct 2025) Minseo Lee, Byeonghyeon Lee, Lucas Yunkyu Lee, Eunsoo Lee, Sangmin Kim, Seunghyeon Song, Joo Chan Lee, Jong Hwan Ko, et al. Paper | Project Page

Architecture

Three-stage progressive pruning pipeline for compact dynamic 3D scenes:

1. Gaussian Sampling — Importance-based retention (opacity x spatial variance x temporal variance)
2. Gaussian Pruning — Opacity thresholding + spatial overlap removal
3. Gaussian Merging — Weighted fusion of similar primitives
4. SVQ Compression — Sub-Vector Quantization with multi-codebook encoding (20.7x compression)

Renderer: gsplat 1.5.3 CUDA kernels (563 FPS @ 50K Gaussians, 800x800)

D-NeRF Benchmark Results

Scene	Test PSNR	Test SSIM	Gaussians (post-prune)
hellwarrior	27.49 dB	0.838	7,944
mutant	23.29 dB	0.910	~100
standup	21.11 dB	0.921	~40K
trex	20.90 dB	0.924	~3K
jumpingjacks	19.93 dB	0.920	~16K
hook	19.71 dB	0.858	19,182
bouncingballs	19.54 dB	0.743	~1K
lego	14.67 dB	0.540	~88K

Primary model: hellwarrior (best quality, 27.49 dB -- near paper target of 28 dB)

Exported Formats

Format	File	Size	Use Case
PyTorch (.pth)	pytorch/huginn_v3.pth	1.9 MB	Training, fine-tuning
SafeTensors	pytorch/huginn_v3.safetensors	1.9 MB	Fast loading, safe
ONNX (projection)	onnx/huginn_projection_v3.onnx	<1 MB	Cross-platform inference
ONNX (SVQ decoder)	onnx/huginn_decoder_v3.onnx	<1 MB	Compressed inference
TensorRT FP16	tensorrt/huginn_projection_v3_fp16.trt	0.5 MB	Edge (Jetson/L4)
TensorRT FP32	tensorrt/huginn_projection_v3_fp32.trt	0.5 MB	Full precision
Per-scene checkpoints	scenes/{scene}_best.pth	varies	Scene-specific models

Usage

import torch
from safetensors.torch import load_file

# Load Gaussian attributes
state = load_file("pytorch/huginn_v3.safetensors")
means = state["means"]        # (N, 3)
quats = state["quats"]        # (N, 4) -- rotation quaternions
scales = state["scales"]      # (N, 3) -- scale factors
sh_coeffs = state["sh_coeffs"]     # (N, 16, 3) -- SH appearance
opacities = state["opacities"]     # (N,)

# Render with gsplat
import gsplat
render_colors, render_alphas, info = gsplat.rasterization(
    means=means, quats=quats, scales=scales,
    opacities=opacities, colors=sh_coeffs,
    viewmats=viewmat, Ks=K, width=800, height=800, sh_degree=3,
)

Training

• Hardware: NVIDIA L4 (23GB VRAM) x 1 GPU
• Framework: PyTorch 2.10 + CUDA 12.8 + gsplat 1.5.3
• Dataset: D-NeRF (8 dynamic scenes, 800x800 native resolution)
• Gaussians: 3M initial, pruned per scene
• Iterations: 30K per scene
• Parallel: 4 scenes simultaneously on 1 GPU
• Total time: ~6 hours for all 8 scenes
• Config: See configs/train_l4_max.yaml

SVQ Compression

Sub-Vector Quantization achieves 20.7x compression (4.8% of original size):

• Position codebook: 256 entries
• Covariance codebook: 256 entries
• Appearance codebook: 256 entries
• Opacity+time codebook: 256 entries

License

Apache 2.0 -- Robot Flow Labs / AIFLOW LABS LIMITED

Built with ANIMA by Robot Flow Labs