Aquarius: A Family of Industry-Level Video Generation Models for Marketing Scenarios

TL;DR

Aquarius addresses the challenge of industrial-scale video generation for marketing by integrating a scalable data pipeline, multi-scale DiT-based diffusion architectures, and memory- and compute-optimized training/inference strategies. The approach combines a 3D VAE backbone with Single DiT, Multimodal DiT, and MoE DiT variants to balance visual fidelity and semantic understanding across 2B–13.4B parameter regimes. Key contributions include Aquarius-Datapipe for large-scale data curation, a bucket-based dataset strategy with Flow Matching training, and sophisticated memory/communication optimization enabling thousands of xPUs to train and deploy diffusion-based video models. Applications span digital humans, video inpainting, and video personalization, demonstrating practical impact for advertising, while future work explores RLHF, MoE diffusion, and unified multimodal modeling to further enhance control, quality, and efficiency.

Abstract

This report introduces Aquarius, a family of industry-level video generation models for marketing scenarios designed for thousands-xPU clusters and models with hundreds of billions of parameters. Leveraging efficient engineering architecture and algorithmic innovation, Aquarius demonstrates exceptional performance in high-fidelity, multi-aspect-ratio, and long-duration video synthesis. By disclosing the framework's design details, we aim to demystify industrial-scale video generation systems and catalyze advancements in the generative video community. The Aquarius framework consists of five components: Distributed Graph and Video Data Processing Pipeline: Manages tens of thousands of CPUs and thousands of xPUs via automated task distribution, enabling efficient video data processing. Additionally, we are about to open-source the entire data processing framework named "Aquarius-Datapipe". Model Architectures for Different Scales: Include a Single-DiT architecture for 2B models and a Multimodal-DiT architecture for 13.4B models, supporting multi-aspect ratios, multi-resolution, and multi-duration video generation. High-Performance infrastructure designed for video generation model training: Incorporating hybrid parallelism and fine-grained memory optimization strategies, this infrastructure achieves 36% MFU at large scale. Multi-xPU Parallel Inference Acceleration: Utilizes diffusion cache and attention optimization to achieve a 2.35x inference speedup. Multiple marketing-scenarios applications: Including image-to-video, text-to-video (avatar), video inpainting and video personalization, among others. More downstream applications and multi-dimensional evaluation metrics will be added in the upcoming version updates.

Figures (14)

• Figure 1: Data Filter Pipeline
• Figure 2: Overview of our 3DVAE.
• Figure 3: Overview of the our VideoDiT architecture.
• Figure 4: Explicit gradient synchronization across tensor-parallel group: preventing parameter drift in image patchify layers.
• Figure 5: Shared Storage Retention: QKV memory deallocation delayed until MLP-in Release.