YODA: Yet Another One-step Diffusion-based Video Compressor

Xingchen Li, Junzhe Zhang, Junqi Shi, Ming Lu, Zhan Ma

Nanjing University

---

📢 Introduction

YODA is a novel neural video codec designed to achieve extreme perceptual quality with efficient inference speed.

While one-step diffusion models have excelled in image compression, applying them to video remains a challenge. YODA overcomes the limitations of traditional methods and existing deep learning baselines by introducing a One-step Diffusion Transformer and a Temporal-Awareness mechanism.

Core Highlights:

• Perceptual Quality: YODA consistently outperforms H.266/VVC and SOTA neural codecs (such as DCVC-RT and PLVC) on perceptual metrics including LPIPS, DISTS, FID, and KID.
• One-Step Denoising: Utilizing a lightweight linear DiT model, YODA performs denoising in a single step, significantly reducing the inference latency associated with diffusion models.
• Temporal-Aware Design: Unlike prior efforts that rely on frozen 2D autoencoders, YODA employs a trainable Temporal-Aware AutoEncoder (TA-AE) to fully exploit inter-frame correlations.

---

🚀 Framework

YODA proposes an end-to-end unified design consisting of three key components:

• Temporal-Aware AutoEncoder (TA-AE): Extracts multiscale features from reference frames to generate a compact latent representation.
• Conditional Latent Coder (CLC): Implicitly models motion within the feature space to perform efficient entropy coding.
• Linear DiT Model: Adopts a linear DiT for efficient one-step denoising.

---

🏆 Performance

YODA demonstrates superior performance across multiple datasets (UVG, HEVC-B, MCL-JCV), surpassing both traditional standards (VTM) and recent neural video codecs (DCVC-RT, DiffVC, PLVC)

Figure: Perceptual quality performance comparisons on UVG, HEVC-B, and MCL-JCV datasets. Lower is better.

---

👁️ Visual Comparison

We provide an interactive video comparison (with sliding view) on our project page to demonstrate the visual reconstruction quality of YODA against the Ground Truth.

---

📂 Data Preparation

We utilized the Vimeo-90K dataset for training and evaluated our model on the UVG, MCL-JCV, and HEVC Class B datasets.

---

🤝 Acknowledgment

We thank the authors of the following projects for their pioneering contributions and open-source efforts:

• DCVC-RT: Towards Practical Real-time Neural Video Compression.
• SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers.
• DC-AE: Deep Compression Autoencoder.