Model reference · open weights
CubeComposer is an open-weight video model from TencentARC. CubeComposer (BF16) weighs 20.0 GB; the smallest configuration that runs it is RTX 3060 12 GB.
Summary of the TencentARC/CubeComposer model card, 2026-10-01
What it is
| Released by | TencentARC |
|---|---|
| Released | 2026-03-03 |
| VRAM | 20.0 GB for the weights |
What it runs on
| Card | Weights | Memory |
|---|---|---|
| RTX 3060 12 GB | tight (encoders offloaded) | 11.6 GB |
| RTX 4060 Ti 16 GB | fits (encoders offloaded) | 15.4 GB |
| RTX 3090 24 GB | tight | 23.4 GB |
| RTX 4090 24 GB | tight | 23.4 GB |
| RTX 5090 32 GB | fits | 31.0 GB |
| L40S 48 GB | fits | 44.0 GB |
| A100 80 GB | fits | 78.2 GB |
| H100 80 GB | fits | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | fits | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | fits | 107 GB |
| H200 141 GB | fits | 138 GB |
| B200 180 GB | fits | 176 GB |
From the model card
TL;DR: Generate one cubemap face per time window with an effective and efficient context mechanism. Then, perspective video becomes 4K 360° without the memory blow‑up or the low‑res‑then‑upscale.
For more details, please visit our project page, paper, and GitHub repo.
We provide two variants of CubeComposer in this repo:
If you find our model helpful in your research, please like this repo, star the GitHub repo and cite:
@article{li2026cubecomposer,
title={CubeComposer: Spatio-Temporal Autoregressive 4K 360° Video Generation from Perspective Video},
author={Li, Lingen and Wang, Guangzhi and Li, Xiaoyu and Zhang, Zhaoyang and Dou, Qi and Gu, Jinwei and Xue, Tianfan and Shan, Ying},
journal={arXiv preprint arXiv:2603.04291},
year={2026}
}
This repository is released under the terms of the LICENSE file.
By cloning, downloading, using, or distributing this repository or any of its models or weights, you agree to comply with the terms and conditions specified in the LICENSE.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works
Running it yourself
Rent a machine by the hour. ComfyUI is installed on it. Open ComfyUI through the tunnel and load the workflow from the model's card on Hugging Face; choose this model's file in its loader.
# on your rented machine (the ssh line is on its page in the console)
# get REPO FILE FOLDER: one file into /workspace/models/FOLDER, where ComfyUI loads it from
get() { hf download "$1" "$2" --local-dir /workspace/hf-files && mkdir -p "/workspace/models/$3" && mv "/workspace/hf-files/$2" "/workspace/models/$3/$4"; }
# the model (9.3 GB)
get TencentARC/CubeComposer cubecomposer-3k/model.safetensors diffusion_models
start-comfyui# on your computer, in a second terminal: ComfyUI in your browser at http://localhost:8188
# HOST and PORT are your machine's, from its page in the console
ssh -L 8188:localhost:8188 dev@HOST -p PORT