💡 The Qwen Image VAE 2.0 technical report presents a high compression Variational Autoencoder suite that improves reconstruction fidelity and diffusability. The problem addressed in this paper is the reconstruction bottleneck of high compression in Variational Autoencoders. To solve this problem, the authors propose an improved architecture featuring Global Skip Connections and expanded latent channels. They also scale training to billions of images and incorporate a synthetic rendering engine to improve performance in text rich scenarios.

The method used in this paper involves implementing an enhanced semantic alignment strategy to make the latent space highly amenable to diffusion modeling. The authors also leverage an asymmetric and attention free encoder decoder backbone to minimize encoding overhead. The performance of Qwen Image VAE 2.0 is evaluated on public reconstruction benchmarks and a new benchmark called OmniDoc TokenBench, which is a collection of real world documents with specialized OCR based evaluation metrics.

The results show that Qwen Image VAE 2.0 achieves state of the art reconstruction performance, demonstrating exceptional capabilities in both general domains and text rich scenarios at high compression ratio. Downstream DiT experiments reveal that the models possess superior diffusability, significantly accelerating convergence compared to existing high compression baselines. Overall, Qwen Image VAE 2.0 establishes itself as a leading model with high compression, superior reconstruction, and exceptional diffusability.

💡 The paper introduces LongLive-2.0, a parallel infrastructure for long video generation that addresses training and inference bottlenecks. The problem with existing methods is that they are slow and require a lot of memory, especially for long videos. To solve this, the authors propose a sequence-parallel autoregressive training method called Balanced SP, which pairs clean-history and noisy-target temporal chunks on each rank, enabling efficient teacher-forcing and reducing GPU memory cost.

The method also uses NVFP4 precision to accelerate GEMM computation during training. Additionally, the authors tune a diffusion model into a long, multi-shot, interactive auto-regressive diffusion model, which can be converted to real-time generation with standalone LoRA weights. For inference, the authors enable W4A4 NVFP4 inference, quantize KV cache into NVFP4 for memory savings, and boost end-to-end throughput with asynchronous streaming VAE decoding.

The results show that LongLive-2.0 achieves up to 2.15x speedup in training and 1.84x in inference. The LongLive-2.0-5B model achieves 45.7 FPS inference while attaining strong performance on benchmarks. The authors claim that LongLive-2.0 is the first NVFP4 training and inference system for long video generation, making it a significant contribution to the field. Overall, the paper presents a novel parallel infrastructure that addresses the speed and memory bottlenecks in long video generation, making it possible to generate high-quality videos in real-time.