We present Magic 1-For-1, an efficient video generation model with optimized memory consumption and inference latency. The key idea is simple: factorize the text-to-video generation task into two separate easier tasks for diffusion step distillation, namely text-to-image generation and image-to-video generation. We verify that with the same optimization algorithm, the overall pipeline is indeed easier to converge.
Magic 1-For-1 uses advanced step distillation algorithm, to train a "generator" model that can produce high-quality videos in just a few steps. This generator model is trained in conjunction with two other models: one that approximates the real data distribution and another that approximates the distribution of the generated data. By aligning these distributions, the generator model learns to produce realistic videos efficiently. We also incorporate CFG distillation to further reduce the computational overhead during inference.
Magic 1-For-1 is an efficient video generation model designed to create short video clips quickly. It simplifies the complex task of generating videos from text by breaking it down into two easier steps: first, generating an image from the text, and then generating a video from that image. This approach leverages the relative simplicity of text-to-image generation. The model incorporates several optimizations, including using multi-modal information to guide the video generation, and streamlining the video generation process to reduce the number of steps required.
The model further enhances efficiency through techniques like adversarial step distillation, which accelerates the generation process, and parameter sparsification, which reduces memory usage. These methods allow the model to generate short video clips in just a few seconds. By using a sliding window approach during inference, the model can also create longer videos, up to a minute in length, while maintaining good visual quality and motion.
@article{yi2025magic,
title={Magic 1-For-1: Generating One Minute Video Clips within One Minute},
author={Hongwei Yi, Shitong Shao, Tian Ye, Jiantong Zhao, Qingyu Yin, Michael Lingelbach, Li Yuan, Yonghong Tian, Enze Xie, Daquan Zhou},
journal={to be updated},
year={2025}
}