MagistrTheOne/ARACHNE-X-ULTRA-VIDEO
ARACHNE-X-ULTRA <div align="center"> </div> <hr> <div align="center"> <a href='https://nullxes.com/arachne-x-ultra'><img src='https://img.shields.io/badge/Project-Page-green'></a> <a href='https://huggingface.co/papers/2510.22200'><img src='https://img.shields.io/badge/Paper-HuggingFace-red'></a> <a href='https://huggingface.co/nullxes/ARACHNE-X-ULTRA'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue'></a> <a href='mailto:ceo@nullxes.com'><img src='https://img.shields.io/badge/Contact-ceo@nullxes.com-lightgrey'></a> </div> <div align="center"> <a href='https://huggingface.co/nullxes/ARACHNE-X-ULTRA/blob/main/LICENSE'><img src='https://img.shields.io/badge/License-MIT-f5de53?&color=f5de53'></a> </div> Model Introduction We introduce ARACHNE-X-ULTRA, a foundational video generation model with 13.6B parameters, delivering strong performance across Text-to-Video, Image-to-Video, and Video-Continuation generation tasks. It particularly excels in efficient, high-quality long video generation — representing NULLXES's first step toward world models. Key Features
🕷️ Unified architecture for multiple tasks: ARACHNE-X-ULTRA unifies Text-to-Video, Image-to-Video, and Video-Continuation within a single framework. One model, natively supports all tasks, consistently strong across each. 🕷️ Long video generation: Natively pretrained on Video-Continuation tasks, enabling minutes-long videos without color drift or quality degradation. 🕷️ Efficient inference: Generates 720p, 30fps video within minutes via a coarse-to-fine generation strategy along temporal and spatial axes. Block Sparse Attention further boosts efficiency at high resolutions. 🕷️ Multi-reward RLHF: Powered by multi-reward Group Relative Policy Optimization (GRPO). Comprehensive evaluations on internal and public benchmarks show performance comparable to leading open-source models and the latest commercial solutions.
For full details, refer to the ARACHNE-X-ULTRA Technical Report.
Quick Start Installation shellgit clone https://github.com/nullxes/ARACHNE-X-ULTRA cd ARACHNE-X-ULTRA shellconda create -n arachne python=3.10 conda activate arachne
pip install torch==2.6.0+cu124 torchvision==0.21.0+cu124 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install ninja psutil packaging flashattn==2.7.4.post1 pip install -r requirements.txt FlashAttention-2 is enabled by default. You can switch to FlashAttention-3 or xformers via model config. Model Download ModelLinkARACHNE-X-ULTRA🤗 Huggingface shellpip install "huggingfacehub[cli]" huggingface-cli download nullxes/ARACHNE-X-ULTRA --local-dir ./weights/ARACHNE-X-ULTRA Inference shell# Text-to-Video (single GPU) torchrun rundemotexttovideo.py --checkpointdir=./weights/ARACHNE-X-ULTRA --enablecompile
Text-to-Video (multi-GPU)
torchrun --nprocpernode=2 rundemotexttovideo.py --contextparallelsize=2 --checkpointdir=./weights/ARACHNE-X-ULTRA --enablecompile
Image-to-Video
torchrun rundemoimagetovideo.py --checkpointdir=./weights/ARACHNE-X-ULTRA --enablecompile
Video-Continuation
torchrun rundemovideocontinuation.py --checkpointdir=./weights/ARACHNE-X-ULTRA --enable_compile
Long Video Generation
torchrun rundemolongvideo.py --checkpointdir=./weights/ARACHNE-X-ULTRA --enable_compile
Interactive Video Generation
torchrun rundemointeractivevideo.py --checkpointdir=./weights/ARACHNE-X-ULTRA --enable_compile
Streamlit UI
streamlit run ./run_streamlit.py --server.fileWatcherType none --server.headless=false
Evaluation Results Text-to-Video (Internal MOS Benchmark) MOS scoreVeo3PixVerse-V5Wan 2.2-T2V-A14BARACHNE-X-ULTRAAccessibilityProprietaryProprietaryOpen SourceOpen SourceArchitecture——MoEDenseTotal Params——28B13.6BActivated Params——14B13.6BText-Alignment ↑3.993.813.703.76Visual Quality ↑3.233.133.263.25Motion Quality ↑3.863.813.783.74Overall Quality ↑3.483.363.353.38 Image-to-Video (Internal MOS Benchmark) MOS scoreSeedance 1.0Hailuo-02Wan 2.2-I2V-A14BARACHNE-X-ULTRAAccessibilityProprietaryProprietaryOpen SourceOpen SourceArchitecture——MoEDenseTotal Params——28B13.6BActivated Params——14B13.6BImage-Alignment ↑4.124.184.184.04Text-Alignment ↑3.703.853.333.49Visual Quality ↑3.223.183.233.27Motion Quality ↑3.773.803.793.59Overall Quality ↑3.353.273.263.17
Community Works Community contributions are welcome. Open a PR or file an Issue to have your work listed.
CacheDiT — Fully cached acceleration for ARACHNE-X-ULTRA via DBCache and TaylorSeer, achieving ~1.7× speedup with minimal precision loss. See their example.
License Model weights are released under the MIT License. All contributions to this repository are licensed under MIT unless otherwise stated. This license does not grant rights to use NULLXES trademarks or patents. See the LICENSE file for full terms.
Usage Considerations ARACHNE-X-ULTRA has not been specifically designed or comprehensively evaluated for every possible downstream application. Developers should account for known limitations of large generative models, carefully assess accuracy, safety, and fairness before deploying in sensitive or high-risk scenarios, and ensure compliance with all applicable laws and regulations — including data protection, privacy, and content safety requirements.
Citation bibtex@misc{nullxes2025arachnexultra, title={ARACHNE-X-ULTRA Technical Report}, author={NULLXES Team}, year={2026}, eprint={2510.22200}, archivePrefix={arXiv}, primaryClass={cs.CV}, }
Contact Questions? Reach out at ceo@nullxes.com