时序与内容协同感知的潜扩散模型用于可控手部图像生成
A Temporal and Content Co-Awareness Latent Diffusion for Controllable Hand Image Generation
提出按去噪时间步与条件复杂度动态调节控制信号注入强度的双驱动调制策略，可迁移到任意「结构条件+外观条件」的可控扩散生成任务（如人体姿态引导生成、姿态迁移），替代固定强度的 ControlNet 式注入。
原文依据 existing methods inject control signals with fixed strength across all timesteps, ignoring the progressive nature of the denoising process
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论可控手部图像生成旨在合成几何准确且外观一致的图像。近年来，扩散模型被广泛应用于手部图像合成。然而，现有方法通过输入级融合或特征级调制注入控制信号，在所有时间步上采用固定强度，忽略了去噪过程的渐进性。本文揭示控制信号的调制取决于去噪状态与条件复杂度。由于语义分布与信息密度存在差异，在这些异构表示之间实现有效交互仍然具有挑战性。为此，我们提出一种时序与内容协同感知的潜扩散方法，引入双驱动调制策略。具体而言，我们设计了一种基于查询的交互机制，以缓解信息冗余并对齐语义分布。借助跨域交互，模型推断所需的控制信息，从而动态调整姿态与外观的注入强度。此外，我们设计了一个姿态不变外观编码器，能够同时捕捉全局外观一致性与局部纹理细节。大量实验验证了我们优于当前最先进方法。
英文原文摘要 1295 字符
Controllable hand image generation aims to synthesize geometrically accurate images with consistent appearance. Recently, diffusion models have been widely applied for hand image synthesis. However, through input-level fusion or feature-level modulation, existing methods inject control signals with fixed strength across all timesteps, ignoring the progressive nature of the denoising process. In this paper, we reveal that the modulation of control signals depends on the denoising state and condition complexity. Due to distinct semantic distributions and information densities, achieving effective interaction among these heterogeneous representations remains challenging. To address this, we propose a Temporal and Content Co-Awareness Latent Diffusion method that introduces a dual-driven modulation strategy. Specifically, we design a query-based interaction mechanism to mitigate information redundancy and align semantic distributions. Leveraging cross-domain interaction, the model infers required control information to dynamically adjust pose and appearance injection strengths. Furthermore, we design a Pose-Invariant Appearance Encoder that captures both global appearance consistency and local texture details. Extensive experiments validate our superiority over state-of-the-art.