文章背景与核心概要
在大型生成模型显著提升图像超分辨率(SR)感知质量的同时,如何保持对低分辨率(LR)观测结果的真实保真度依然是一个持续存在的难题。本文指出,基于隐空间表征构建的扩散Transformer(DiT)存在一个关键局限性:变分自编码器(VAE)的压缩瓶颈削弱了细粒度空间信息,导致模型产生缺乏输入图像依据的“幻觉”细节。
为了解决这一问题,作者引入了像素基础超分辨率(Pixel-Grounded Super-Resolution, PGSR)框架。通过提取并重用VAE编码前的像素证据,PGSR在无需重新训练庞大预训练骨干网络的情况下,实现了更出色的真实感与保真度平衡。该研究为基于扩散的图像恢复任务提供了轻量且高效的改进思路。
当隐空间遗忘像素:恢复扩散Transformer超分辨率中的保真度
作者: Yu Shi, Yuyao Zhang, Yu-wing Tai
机构 / 学科领域: 计算机视觉与模式识别 (cs.CV)、人工智能 (cs.AI)
arXiv ID: arXiv:2608.09133 [cs.CV]
提交时间: 2026年8月10日
📌 执行摘要
While large generative models have significantly improved the perceptual quality of image super-resolution (SR), maintaining true fidelity to low-resolution (LR) observations remains a persistent hurdle.
This paper identifies a critical limitation in Diffusion Transformers (DiTs) built on latent representations: the compression bottleneck of the Variational Autoencoder (VAE) weakens fine-grained spatial information, leading to hallucinated details that lack grounding in the input image. To address this, the authors introduce the Pixel-Grounded Super-Resolution (PGSR) framework. By extracting and reusing pre-VAE pixel evidence, PGSR achieves a superior realism-fidelity trade-off without retraining massive pretrained backbones.
虽然大型生成模型显著提升了图像超分辨率(SR)的感知质量,但保持对低分辨率(LR)观测结果的真正保真度依然是一个长期的障碍。本文指出了基于隐空间表征构建的扩散Transformer(DiT)的一个关键局限性:变分自编码器(VAE)的压缩瓶颈削弱了细粒度空间信息,导致模型产生缺乏输入图像依据的幻觉细节。为了解决这一问题,作者引入了像素基础超分辨率(PGSR)框架。通过提取并重用VAE编码前的像素证据,PGSR在无需重新训练庞大预训练骨干网络的前提下,实现了更优的真实感与保真度权衡。
🔍 核心洞察与隐空间扩散的问题所在
- The Compression Bottleneck: Standard latent diffusion models rely on VAEs to compress images into latent space. During this process, high-frequency spatial details are often lost, causing the model to "forget" precise pixel configurations.
- Weakly Grounded Hallucinations: As a result, subsequent generation steps introduce fine details that look visually appealing but do not accurately reflect the original LR observation.
- 压缩瓶颈: 标准的隐空间扩散模型依赖VAE将图像压缩到隐空间中。在此过程中,高频空间细节往往会丢失,导致模型“遗忘”精确的像素配置。
- 弱基础的幻觉: 因此,后续的生成步骤会引入一些看起来视觉效果不错、但并未准确反映原始LR观测结果的精细细节。
🛠️ 提出方法:像素基础超分辨率 (PGSR)
The PGSR framework re-evaluates generative SR from a representation perspective. Instead of relying exclusively on compressed latent conditions, it preserves pre-VAE pixel evidence from the upsampled LR image and integrates it across two crucial stages:
- Condition-Side Trajectory Guidance: Fuses LR-derived pixel evidence directly with the latent LR condition to steer the latent restoration trajectory.
- Decoder-Side Pixel Grounding: Injects multi-scale pixel features straight into the frozen VAE decoder, firmly grounding the final image rendering in authentic, LR-observed cues.
Efficiency & Scalability
To make adaptation practical for large pretrained DiT models: * The core latent autoencoder and the main flow-matching backbone are kept completely frozen. * Only lightweight restoration modules are trained. * An efficient local-window attention variant is introduced to boost high-resolution efficiency and scalability.
PGSR框架从表征视角重新评估了生成式超分辨率。它没有单纯依赖压缩后的隐空间条件,而是保留了来自上采样LR图像的VAE编码前像素证据,并将其整合到两个关键阶段中:
- 条件侧轨迹引导: 将LR导出的像素 evidence(证据)直接与隐空间LR条件相融合,以引导隐空间恢复轨迹。
- 解码器侧像素基础化: 将多尺度像素特征直接注入到冻结的VAE解码器中,将最终的图像渲染牢牢锚定在真实的LR观测线索上。
效率与可扩展性
为了使这种适配方案对大型预训练DiT模型切实可行: * 核心的隐空间自编码器和主要的流匹配(flow-matching)骨干网络保持完全冻结。 * 仅训练轻量级的恢复模块。 * 引入了高效的局部窗口注意力变体,以提升高分辨率下的效率和可扩展性。
📊 结果与性能
Extensive empirical evaluations show that PGSR effectively balances realism and fidelity, consistently outperforming existing latent generative SR methods by producing sharper, more visually convincing, and faithful results.
广泛的实证评估表明,PGSR 有效地平衡了真实感与保真度,通过生成更锐利、视觉上更具说服力且忠于原图的结果,始终优于现有的隐空间生成式超分辨率方法。
🔗 链接与资源
- Paper: arXiv:2608.09133
- PDF Download: Direct PDF Link
- HTML Version: arXiv HTML (Experimental)
- Citations & Metrics: Google Scholar | Semantic Scholar
- 论文地址: arXiv:2608.09133
- PDF 下载: PDF 直链
- HTML 版本: arXiv HTML (实验性)
- 引用与指标: Google Scholar | Semantic Scholar