---
title: "FlashDecoder: Real-Time Latent-to-Pixel Streaming Video Decoder"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.14898v1"
permalink: "https://feed.astralor.com/items/itm_9f624e4fdce3348d3f3e7b78"
published: "2026-07-16T12:16:12.000Z"
score: 76
channel: "tech"
tags: ["FlashDecoder","Transformer","Wan2.1","Wan2.2","Real-time Generation","Video Decoding"]
---

# FlashDecoder: Real-Time Latent-to-Pixel Streaming Video Decoder

## AI commentary

> The highlight of this paper is replacing the expensive 3D convolutional decoder in video diffusion models with a pure Transformer, achieving comparable quality and several times speedup. This is a highly practical optimization for engineers working on video generation, real-time rendering, and model deployment. The streaming inference design—constant latency regardless of video length—paves the way for live-level video generation.

## At a glance

FlashDecoder is a pure-Transformer video decoder designed for real-time video generation. It decodes latents to pixels frame by frame, with each frame attending only to a fixed-size window of past frames via a rolling KV cache, ensuring constant latency and bounded memory regardless of video length. On Wan2.1 and Wan2.2 latent spaces, it achieves PSNR of 41.55dB at 1080p—comparable to convolutional decoders—while decoding 3.6–4.7× faster and using up to 11× less memory. With architecture-aware optimizations, speedup reaches 12×.

## Key points

- Frame-by-frame streaming decoding with rolling KV cache for fixed-window attention, achieving latency independent of video length.
- Matches convolutional decoder reconstruction quality at 1080p (PSNR 41.55dB vs 41.49dB).
- Decodes 3.6–4.7× faster with up to 11× less memory on a single H100 GPU.
- Architecture-aware inference optimizations widen speedup to 12×.

## Full text

Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_9f624e4fdce3348d3f3e7b78