语义先行：用异步潜扩散协调语义与纹理建模
Semantics Lead the Way: Harmonizing Semantic and Texture Modeling with Asynchronous Latent Diffusion
把预训练视觉编码器的语义潜变量与 VAE 纹理潜变量拼接，并用带时间偏移的两套噪声调度异步去噪，让语义先于纹理成形；这一「语义先行」的调度可直接嵌入现有 DiT/LDM 训练流程，用于加速收敛并提升 FID。
原文依据 denoise the semantic and texture latents asynchronously using separate noise schedules: semantics precede textures by a temporal offset
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论潜扩散模型（LDM）本身遵循由粗到细的生成过程，其中高层语义结构的生成略早于细粒度纹理。这表明先行的语义有可能通过提供语义锚点来助益纹理生成。近期进展已将预训练视觉编码器的语义先验引入以进一步增强 LDM，但它们仍同步地对语义与 VAE 编码的纹理进行去噪，忽略了上述次序。基于这些观察，我们提出语义先行扩散（SFD），一种显式优先形成语义的潜扩散范式。SFD 首先通过一个专用的语义 VAE 从预训练视觉编码器中提取紧凑的语义潜变量，并将其与纹理潜变量组合成复合潜变量。SFD 的核心是用各自独立的噪声调度对语义潜变量与纹理潜变量进行异步去噪：语义以时间偏移领先于纹理，从而为纹理细化提供更清晰的高层引导，并实现自然的由粗到细生成。在带引导的 ImageNet 256x256 上，SFD 取得弗雷歇初始距离 1.06（LightningDiT-XL）与 1.04（1.0B LightningDiT-XXL），同时相比无引导的原始 DiT 实现最高 100 倍的收敛加速。SFD 还改进了 ReDi 与 VA-VAE 等现有方法，验证了异步、语义主导建模的有效性。
英文原文摘要 1408 字符
Latent Diffusion Models (LDMs) inherently follow a coarse-to-fine generation process, where high-level semantic structure is generated slightly earlier than fine-grained texture. This indicates the preceding semantics potentially benefit the texture generation by providing a semantic anchor. Recent advances have integrated semantic priors from pretrained visual encoders to further enhance LDMs, yet they still denoise semantic and VAE-encoded texture synchronously, neglecting such ordering. Observing these, we propose Semantic-First Diffusion (SFD), a latent diffusion paradigm that explicitly prioritizes semantic formation. SFD first constructs composite latents by combining the compact semantic latent, which is extracted from pretrained visual encoder via a dedicated Semantic VAE, with the texture latent. The core of SFD is to denoise the semantic and texture latents asynchronously using separate noise schedules: semantics precede textures by a temporal offset, providing clearer high-level guidance for texture refinement and enabling natural coarse-to-fine generation. On ImageNet 256x256 with guidance, SFD achieves FID 1.06 (LightningDiT-XL) and FID 1.04 (1.0B LightningDiT-XXL), while achieving up to 100xfaster convergence than original DiT without guidance. SFD also improves existing methods like ReDi and VA-VAE, demonstrating the effectiveness of asynchronous, semantics-led modeling.