STORM-B/8

Closed weights University of Southern California,Georgia Institute of Technology,Stanford University,NVIDIA 100.6M parameters December 2024

No estimate

No hardware requirements for this model

The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation
University of Southern California,Georgia Institute of Technology,Stanford University,NVIDIA
Organisation type
Academia,Academia,Academia,Industry
Country
United States of America
Published
31 December 2024
Authors
Jiawei Yang, Jiahui Huang, Yuxiao Chen, Yan Wang, Boyi Li, Yurong You, Apoorva Sharma, Maximilian Igl, Peter Karkus, Danfei Xu, Boris Ivanovic, Yue Wang, Marco Pavone

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
3D modeling
Task
3D reconstruction, 3D segmentation
Numerical format
TF32

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters
100.6M

A.3 Baseline Implementations: "The number of trainable parameters of these models are controlled to be similar, i.e., GS-LRM has 86.68M parameters, LGM has 103.29M parameters, while our default STORM has 100.60M parameters. 3.3 Implementation, Model architecture: "By default, we use a 12-layer Vision Transformer (ViT-B) (Dosovitskiy, 2020) with full attention and a patch size of 8, along with M = 16 motion tokens." https://github.com/NVlabs/GaussianSTORM?tab=readme-ov-file#training Training: "…

Training data
tokens

4 Experiments, Datasets: "We primarily conduct experiments on the Waymo Open Dataset (Sun et al., 2020), which contains 1,000 sequences of driving logs: 798 sequences for training and 202 for validation. Each sequence consists of a 20-second video recorded at 10FPS." 10 frames/s * 20 s/scene = 200 frames/scene 200 frames/scene * 798 scene = 159600 frames (training) scene range: [0.1,20] s, or [1,200] frames Sampling scheme: 3.3 Implementation, Supervision and loss functions: "During traini…

Batch size
64

Global batch size is 64; repo suggests per-GPU split of 4

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware
NVIDIA A100
Chips used
16
Power draw
12.6 kW

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Why it is tracked
SOTA improvement

Abstract: "Extensive experiments on public datasets show that STORM achieves precise dynamic scene reconstruction, surpassing state-of-the-art per-scene optimization methods (+4.3 to 6.6 PSNR) and existing feed-forward approaches (+2.1 to 4.7 PSNR) in dynamic regions. STORM reconstructs large-scale outdoor scenes in 200ms, supports real-time rendering, and outperforms competitors in scene flow estimation, improving 3D EPE by 0.422m and Acc5 by 28.02%." See Tables 1,2,3 - comparisons to EmerNeRF…

Record confidence
Confident
Citations
38

Sources

Where this record came from and when it was last checked.

Reference
STORM: Spatio-Temporal Reconstruction Model for Large-Scale Outdoor Scenes
Last updated
25 May 2026

What the numbers mean

Where it came from

STORM-B/8 was published by University of Southern California,Georgia Institute of Technology,Stanford University,NVIDIA, in United States of America, in December 2024. academia,Academia,Academia,Industry is the category the publisher falls under.

It works in 3D modeling, and is recorded as doing 3D reconstruction, 3D segmentation.

Its weights were never published, so it can only be reached through its provider. No graphics card changes that.

Training and provenance

It is tracked in the underlying dataset for one reason in particular: sOTA improvement.

Answers

STORM-B/8 — common questions

01

Is STORM-B/8 open source?

The licensing for STORM-B/8 was never recorded in our source data. We treat unstated licensing as closed, because an unrecorded licence is not one to rely on.

02

How many parameters does STORM-B/8 have?

STORM-B/8 has 100.6M parameters. A.3 Baseline Implementations: "The number of trainable parameters of these models are controlled to be similar, i.e., GS-LRM has 86.68M parameters, LGM has 103.29M parameters, while our default STORM has 100.60M parameters. 3.3 Implementation, Model architecture: "By default, we use a 12-layer Vision Transformer (ViT-B) (Dosovitskiy, 2020) with full attention and a patch size of 8, along with M = 16 motion tokens." https://github.com/NVlabs/GaussianSTORM?tab=readme-ov-file#training Training: "Multi-GPU example that reproduces the paper's STORM-B/8 model: https://github.com/NVlabs/GaussianSTORM/blob/main/storm/models/storm.py def STORM_B_8(**kwargs): return STORM(patch_size=8, embed_dim=768, depth=12, num_heads=12, **kwargs) def STORM_L_8(**kwargs): return STORM(patch_size=8, embed_dim=1024, depth=24, num_heads=16, **kwargs) def STORM_B_16(**kwargs): return STORM(patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs) def STORM_L_16(**kwargs): return STORM(patch_size=16, embed_dim=1024, depth=24, num_heads=16, **kwargs) def STORM_XL_8(**kwargs): return STORM(patch_size=8, embed_dim=1152, depth=28, num_heads=16, **kwargs) def STORM_H_8(**kwargs): return STORM(patch_size=8, embed_dim=1280, depth=32, num_heads=16, **kwargs) def STORM_H_16(**kwargs): return STORM(patch_size=16, embed_dim=1280, depth=32, num_heads=16, **kwargs) PyTorch model.parameters(): with "dummy decoder": STORM-B/8: 100.60M (100,598,707) STORM-L/8: 327.08M (327,081,139) STORM-XL/8: 476.55M (476,547,379) STORM-H/8: 665.91M (665,905,587) STORM-B/16: 103.42M (103,417,907) STORM-L/16: 330.82M (330,817,331) STORM-H/16: 670.56M (670,558,771) with "conv decoder", i.e. "Latent-STORM" (adds ~30M): STORM-B/8: 133.23M (133,230,019) STORM-L/8: 359.13M (359,125,699) STORM-XL/8: 508.30M (508,298,563) STORM-H/8: 697.36M (697,363,395) STORM-B/16: 134.21M (134,211,523) STORM-L/16: 360.43M (360,434,371) STORM-H/16: 699.00M (698,999,235). That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

03

Who created STORM-B/8?

STORM-B/8 was published by University of Southern California,Georgia Institute of Technology,Stanford University,NVIDIA, based in United States of America, categorised as academia,Academia,Academia,Industry.

04

When was STORM-B/8 released?

STORM-B/8 was published in December 2024. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

05

What is STORM-B/8 used for?

STORM-B/8 works in 3D modeling, and is recorded as handling 3D reconstruction, 3D segmentation. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

06

What GPU do I need to run STORM-B/8?

None. STORM-B/8 is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.

Source

Original publication

Record last updated 25 May 2026

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.