Infinity⭐️: Unified Spacetime AutoRegressive Modeling for Visual Generation

• Nov 7, 2025: 🔥 Paper, Training and Inference Codes && Checkpoints && Demo Website released!
• Sep 18, 2025: 🎉 InfinityStar is accepted as NeurIPS 2025 Oral.

We provide a demo website for you to play with InfinityStar and generate videos. Enjoy the fun of bitwise video autoregressive modeling!

We introduce InfinityStar, a unified spacetime autoregressive framework for high-resolution image and dynamic video synthesis.

• 🧠 Unified Spacetime Model: A purely discrete, autoregressive approach that jointly captures spatial and temporal dependencies within a single, elegant architecture.

• 🎬 Versatile Generation: This unified design naturally supports a variety of generation tasks such as text-to-image, text-to-video, image-to-video, and long interactive video synthesis via straightforward temporal autoregression.

• 🏆 Leading Performance & Speed: Through extensive experiments, InfinityStar scores 83.74 on VBench, outperforming all autoregressive models by large margins, even surpassing diffusion competitors like HunyuanVideo, approximately 10x faster than leading diffusion-based methods.

• 📖 Pioneering High-Resolution Autoregressive Generation: To our knowledge, InfinityStar is the first discrete autoregressive video generator capable of producing industrial-level 720p videos, setting a new standard for quality in its class.

• Training Code
• Web Demo
• InfinityStar Inference Code
• InfinityStar Models Checkpoints
• InfinityStar-Interact Inference Code
• InfinityStar-Interact Checkpoints

1. We use FlexAttention to speedup training, which requires torch>=2.5.1.
2. Install other pip packages via pip3 install -r requirements.txt.

We provide a comprehensive workflow for training and finetuning our model, covering data organization, feature extraction, and training scripts. For detailed instructions, please refer to data/README.md.

• 720p Video Generation: Use tools/infer_video_720p.py to generate 5-second videos at 720p resolution. Due to the high computational cost of training, our released 720p model is trained for 5-second video generation. This script also supports image-to-video generation by specifying an image path.

  python3 tools/infer_video_720p.py

• 480p Variable-Length Video Generation: We also provide an intermediate checkpoint for 480p resolution, capable of generating videos of 5 and 10 seconds. Since this model is not specifically optimized for Text-to-Video (T2V), we recommend using the experimental Image-to-Video (I2V) and Video-to-Video (V2V) modes for better results. To specify the video duration, you can edit the generation_duration variable in tools/infer_video_480p.py to either 5 or 10. This script also supports image-to-video and video continuation by providing a path to an image or a video.

  python3 tools/infer_video_480p.py

• 480p Long Interactive Video Generation: Use tools/infer_interact_480p.py to generate a long interactive video in 480p. This script supports interactive video generation. You can provide a reference video and multiple prompts. The model will generate a video interactively with your assistance.

  python3 tools/infer_interact_480p.py

If our work assists your research, feel free to give us a star ⭐ or cite us using:

@Article{VAR, title={Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction}, author={Keyu Tian and Yi Jiang and Zehuan Yuan and Bingyue Peng and Liwei Wang}, year={2024}, eprint={2404.02905}, archivePrefix={arXiv}, primaryClass={cs.CV} } 

@misc{Infinity, title={Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis}, author={Jian Han and Jinlai Liu and Yi Jiang and Bin Yan and Yuqi Zhang and Zehuan Yuan and Bingyue Peng and Xiaobing Liu}, year={2024}, eprint={2412.04431}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2412.04431}, } 

@misc{InfinityStar, title={InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation}, author={Jinlai Liu and Jian Han and Bin Yan and Hui Wu and Fengda Zhu and Xing Wang and Yi Jiang and Bingyue Peng and Zehuan Yuan}, year={2025}, eprint={2511.04675}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2511.04675}, } 

This project is licensed under the MIT License - see the LICENSE file for details.