Waver 1.0 - An AI video generation model launched by ByteDance

What is Waver 1.0?

Waver 1.0 is a next-generation video generation model launched by ByteDance. Based on the Modified Stream Transformer architecture, it supports text-to-video (T2V), image-to-video (I2V), and text-to-image (T2I) generation, all within a single framework without the need to switch models. It supports resolutions up to 1080p and flexible video lengths from 2 to 10 seconds, excelling at capturing complex motion, and producing videos with excellent performance in motion amplitude and temporal consistency. On the Waver-Bench 1.0 and Hermes motion test suites, Waver 1.0 outperforms existing open-source and closed-source models. It supports video generation in various artistic styles, including ultra-realistic, animated, claymation, and plush.

Main features of Waver 1.0

• Integrated generationIt supports text-to-video (T2V), image-to-video (I2V), and text-to-image (T2I) generation without switching models.
• High resolution and flexible lengthSupports resolutions up to 1080p and video lengths of 2-10 seconds, with flexible adjustment of resolution and aspect ratio.
• Complex motion modelingIt excels at capturing complex motion, and the generated videos perform exceptionally well in terms of motion amplitude and temporal consistency.
• Multi-camera narrativeIt can generate multi-shot, coherent narrative videos, ensuring a high degree of consistency in core themes, visual style, and overall atmosphere.
• Artistic style supportSupports video generation in various art styles, such as ultra-realistic, animation, clay, and plush.
• performance advantagesOn Waver-Bench 1.0 and the Hermes motion test suite, it outperforms existing open-source and closed-source models.
• Inference optimization: APG technology is used to reduce artifacts and improve the realism of the generated video.
• Training strategyThe motion generation capability is optimized by training with low-resolution videos and gradually increasing the resolution.
• Tip labelBy using prompt labels to distinguish different types of training data, the generation effect can be improved.

Technical principles of Waver 1.0

• Model ArchitectureWaver 1.0 employs a Hybrid Stream DiT architecture, using Wan-VAE to obtain latent variables from compressed video, and flan-t5-xxl and Qwen2.5-32B-Instruct to extract text features. A DiT model is built based on a modified stream Transformer. Video and text modalities are fused using a dual-stream + single-stream approach.
• 1080p generationWaver-Refiner employs a DiT architecture and is trained using a stream matching method. It first upsamples low-resolution videos (480p or 720p) to 1080p, then adds noise. Using this noisy low-resolution video as input, it outputs a high-quality 1080p video. A window attention mechanism is used to reduce inference steps, significantly improving inference speed.
• Training methodsTraining on low-resolution videos is crucial for learning motion. Dedicate significant computational resources to training on 192p videos, then gradually increase the resolution to 480p and 720p. Follow the stream-matching training settings in SD3, gradually increasing the sigma shift value during training on 480p and 720p videos.
• Tip labelA cue labeling method is used to distinguish different types of training data, assigning different labels to training data based on video style and video quality. During training, cues describing video style are added before the title, and cues describing video quality are added to the end of the training title. During inference, cues describing poor quality (such as low resolution or slow motion) are included as negative cues.
• Inference optimizationAPG extends APG to video generation to enhance realism and reduce artifacts. APG decomposes the update term in CFG into parallel and orthogonal components and reduces the weight of the parallel component to achieve high-quality generation without oversaturation.

Waver 1.0 project address

• Project official websitehttp://www.waver.video/
• Github repositoryhttps://github.com/FoundationVision/Waver
• arXiv technical paper: https://arxiv.org/pdf/2508.15761

Application scenarios of Waver 1.0

• Content creationIt can be used for story creation, advertising, short dramas, etc., to transform text descriptions into vivid videos.
• Product ShowcaseIt can convert product images into dynamic display videos, suitable for e-commerce live streaming, product display, and virtual try-on.
• Education and TrainingTeaching content or training documents can be transformed into interactive videos to enhance the learning experience.
• social mediaQuickly generate video content suitable for social media posting to attract user attention.
• Animation ProductionConvert static images into animations; suitable for animation production, character-driven animation, and special effects generation.
• Game developmentGenerate dynamic scenes and character animations for the game, enhancing the game's immersive experience.