HADD: A hierarchy-aware disentangled diffusion framework with spatio-temporal denoising for 3D human pose estimation Jinze Jiang, Zhongheng Jian, Hualing Zheng, Minsi Liang, Yuxing Shi, Liming Zhang, Gennady S Cymbalyuk Abstract Monocular 3D Human Pose Estimation (3D HPE) is a core task in computer vision, which provides critical technical support for applications including VR/AR, immersive human-computer interaction, digital human animation, as well as sports motion capture, athletic technique diagnosis, physical fitness assessment and sports rehabilitation training. However, existing methods still face three key limitations: diffusion-based methods lack explicit human anatomical prior constraints in the diffusion process; disentangled-based methods suffer from severe hierarchical error accumulation along the skeletal kinematic tree; and most mainstream methods fail to fully mine the fine-grained hierarchical spatio-temporal correlations of the human skeleton, leading to increased estimation error for high-hierarchy joints. To address these challenges, we propose a novel Hierarchy-Aware Disentangled Diffusion framework with Spatio-Temporal Denoising (HADD) for monocular 3D HPE, which deeply integrates disentanglement strategy with the diffusion model and embeds skeletal hierarchical information into the full diffusion pipeline. In the forward diffusion process, HADD disentangles 3D pose into bone length and bone direction features, and injects Gaussian noise into the two features separately. In the reverse denoising process, we design a Hierarchical Spatial and Temporal Denoising (HSTD) module to accurately model the hierarchical spatio-temporal relationships of the human skeleton and alleviate error amplification of high-hierarchy joints. Meanwhile, a hybrid loss function combining 3D disentanglement loss and 3D pose loss is constructed to realize dual supervision at the bone and joint levels. Extensive experiments on Human3.6M and MPI-INF-3DHP datasets show that HADD achieves consistent performance improvement over state-of-the-art methods. On Human3.6M, it reduces average MPJPE by up to 12.9% compared with traditional disentangled-based methods, and outperforms mainstream diffusion-based probabilistic methods under lightweight inference configurations. On MPI-INF-3DHP, HADD achieves an MPJPE of 29.2 mm, a PCK of 98.5%, and an AUC of 78.1 under single-hypothesis inference, reaching new state-of-the-art performance on this dataset under the single-hypothesis inference setting with only ground-truth 2D joint sequences as input. This work provides an effective solution to the core limitations of existing 3D HPE methods, and also offers reliable technical support for fine-grained motion analysis and quantitative assessment in sports training and rehabilitation scenarios. 1. Introduction 3D Human Pose Estimation (3D HPE) is a core research direction in the field of computer vision, with its primary goal being to accurately regress the 3D coordinates of human joints in 3D space from 2D human pose sequences input via monocular vision. This technology provides critical technical support for a wide spectrum of practical applications, including virtual and augmented reality (VR/AR) [], immersive human-computer interaction [], digital human animation [], intelligent video surveillance [], as well as competitive sports motion capture, athletic technique diagnosis and optimization and public physical fitness assessment. In these scenarios, the precise capture and analysis of human 3D skeletal motion can realize realistic real-time driving of virtual avatars, construction of natural non-contact interactive systems, fine-grained recognition of complex human actions, quantitative assessment of motor function in rehabilitation scenarios, and standardized quantitative evaluation of sports technical movements. Especially in competitive sports training and youth fitness testing, 3D pose estimation breaks through the limitations of traditional manual observation and 2D plane analysis, enabling three-dimensional refined analysis of dynamic movements such as gymnastics, track and field, racket sports and martial arts, assisting coaches in accurately identifying technical defects, formulating personalized training schemes, and effectively reducing the incidence of sports injuries. Current mainstream 3D HPE methods [,] generally adopt a two-stage processing strategy: (1) first, high-precision 2D joint coordinates are detected from raw RGB images using mature 2D human pose estimation algorithms; (2) then, a 2D-to-3D lifting operation is performed to compensate for the lack of depth dimension information, ultimately yielding the estimation results of 3D human poses. In recent years, monocular 3D HPE technology has witnessed rapid development, and researchers have proposed a variety of solutions to address the core challenge of depth ambiguity, which can be mainly divided into two research veins. (1) The first is the spatiotemporal information mining direction: such methods leverage Convolutional Neural Networks (CNNs) [] or Transformer architectures [] to fully explore temporal information and spatial context information in video sequences, thereby compensating for the depth information loss in the 2D-to-3D mapping. For example, Bian et al. [] proposed a novel human pose estimation framework that adopts a task-specific backbone network and an uncertainty-aware regression head, integrating fine-grained local details and global contextual information to improve the accuracy and robustness of pose estimation. These methods significantly improve the continuity and accuracy of pose estimation through the modeling of sequence information. (2) The second is the human pose prior introduction direction: such methods [,] constrain the generation space of 3D poses by incorporating prior knowledge such as human anatomical structures and pose distributions, alleviating the estimation uncertainty caused by depth ambiguity. Among them, diffusion model-based methods [,] have achieved the current State-of-the-Art (SOTA) performance in this field by directly regressing 3D joint coordinates. For example, Shi et al. [] proposed the motion-semantics-based diffusion (MS-Diff) framework for monocular video-based 3D human pose estimation, which integrates high-level motion semantics via a Multimodal Diffusion Interaction (MDI) module and applies spectral feature regularization through a Spectral Convolutional Regularization (SCR) module to reduce noise, resolve pose ambiguities, and handle occlusions effectively. And disentangled-based methods [,] have taken a different approach by decomposing the 3D pose estimation task into subtasks of bone length and bone direction prediction, and explicitly incorporating human pose prior constraints based on the forward kinematic structure of the human skeleton. For example, Liu et al. [] proposed a novel framework that achieves viewpoint invariance for 3D human pose estimation by explicitly disentangling motion and view features through a lightweight view estimator and a view-aware inference pipeline. In recent years, monocular 3D HPE technology has witnessed rapid development, and researchers have proposed a variety of solutions to address the core challenge of depth ambiguity, which can be mainly divided into two research veins. (1) The first is the spatiotemporal information mining direction: such methods leverage Convolutional Neural Networks (CNNs) [] or Transformer architectures [] to fully explore temporal information and spatial context information in video sequences, thereby compensating for the depth information loss in the 2D-to-3D mapping. For example, Bian et al. [] proposed a novel human pose estimation framework that adopts a task-specific backbone network and an uncertainty-aware regression head, integrating fine-grained local details and global contextual information to improve the accuracy and robustness of pose estimation. These methods significantly improve the continuity and accuracy of pose estimation through the modeling of sequence information, and can better adapt to the analysis requirements of continuous high-dynamic movements in sports scenarios. (2) The second is the human pose prior introduction direction: such methods [,] constrain the generation space of 3D poses by incorporating prior knowledge such as human anatomical structures and pose distributions, alleviating the estimation uncertainty caused by depth ambiguity. Among them, diffusion model-based methods [,] have achieved the current State-of-the-Art (SOTA) performance in this field by directly regressing 3D joint coordinates. For example, Shi et al. [] proposed the motion-semantics-based diffusion (MS-Diff) framework for monocular video-based 3D human pose estimation, which integrates high-level motion semantics via a Multimodal Diffusion Interaction (MDI) module and applies spectral feature regularization through a Spectral Convolutional Regularization (SCR) module to reduce noise, resolve pose ambiguities, and handle occlusions effectively. And disentangled-based methods [,] have taken a different approach by decomposing the 3D pose estimation task into subtasks of bone length and bone direction prediction, and explicitly incorporating human pose prior constraints based on the forward kinematic structure of the human skeleton. This idea is highly consistent with the kinematic analysis logic in sports biomechanics, and has natural application potential in the field of sports technology analysis. For example, Liu et al. [] proposed a novel framework that achieves viewpoint invariance for 3D human pose estimation by explicitly disentangling motion and view features through a lightweight view estimator and a view-aware inference pipeline. Challenges. Despite the remarkable progress made by existing methods, both diffusion model-based methods and disentangled-based methods still have their own limitations, and neither of the two types of methods has fully explored the hierarchical structural information of the human skeleton, resulting in an insurmountable bottleneck in model performance. Specifically, there are three core challenges: (1) Existing diffusion model-based methods [–] directly add Gaussian noise to the original 3D pose as a whole without performing structural decomposition on the 3D pose, making it impossible to effectively learn explicit prior knowledge such as the temporal consistency of bone length and the variation rules of joint angles in the human skeleton. The learning process of the model lacks the constraints of human body structure, resulting in an insufficiently refined modeling of 3D poses; (2) Although disentangled-based methods [–] can explicitly incorporate the anatomical priors of the human skeleton through the decomposition of bone length and bone direction, affected by the tree-like hierarchical structure of the human skeleton, the prediction errors of bone length and bone direction will continuously propagate and amplify along the skeletal hierarchy, forming a severe hierarchical error accumulation effect. This ultimately leads to their overall performance being far lower than that of mainstream diffusion model-based methods, failing to exert the practical effect of human pose priors. In sports scenarios, this error accumulation will be more prominent in terminal joints such as hands and ankles that are critical for technical movement evaluation, directly affecting the reliability of sports motion analysis; (3) Whether it is Transformer-based spatiotemporal modeling methods [,], or diffusion model-based and disentangled-based 3D pose estimation methods [–], all ignore the fine-grained hierarchical correlations of the human skeleton. The human skeleton can be divided into multiple hierarchies according to the depth of the kinematic tree, and the spatial position and temporal variation of each joint are strongly correlated with its parent and child joints. With the increase of joint hierarchy, the estimation error of existing methods rises significantly, and hierarchical error accumulation has become a key factor restricting model performance. To address the above core issues, we propose a Hierarchy-Aware Disentangled Diffusion Framework with Spatio-Temporal Denoising (HADD), which innovatively combines the disentanglement strategy with the diffusion model and deeply integrates the hierarchical information of the human skeleton into the forward and reverse processes of the diffusion model. HADD realizes the dual optimization of structured prior learning and hierarchical spatiotemporal modeling. The core design ideas of HADD are as follows: (1) on the one hand, the 3D pose is disentangled into bone length and bone direction in the forward process of the diffusion model, and we add noise to the disentangled bone length and bone direction respectively, allowing the model to focus on the temporal consistency of bone length and the spatial variation rules of bone direction separately. HADD can learn the explicit priors of human pose more efficiently, while avoiding the hierarchical error accumulation caused by direct disentangled prediction; (2) on the other hand, in the reverse denoising process of the diffusion model, we design a targeted hierarchical spatiotemporal denoiser to strengthen the spatial and temporal correlations between each joint and its parent and child joints, accurately modeling the hierarchical structural information of the human skeleton and alleviating the estimation error of higher-hierarchy joints. The contributions of our work can be summarized as the following three points: - We deeply integrate disentangled representation strategy with the diffusion model, and fully embed the hierarchical structural information of the human skeleton into the full pipeline of the diffusion model. This design alleviates two core limitations of existing methods, namely the lack of explicit human anatomical prior constraints in diffusion-based methods and severe hierarchical error accumulation in disentangled-based methods. - Based on the hierarchical kinematic structure of the human skeleton, we disentangle the 3D pose into two independent feature dimensions: bone length and bone direction. And we add Gaussian noise to each of them in the diffusion forward process. Meanwhile, we design a disentanglement loss function to guide model training, which enables the model to efficiently learn the explicit prior knowledge of human pose. - For the reverse denoising process, we design a Hierarchical Spatial and Temporal Denoising (HSTD) module, which consists of HSDM and HTDM sub-modules to accurately model the hierarchical spatio-temporal relationships of the human skeleton and mitigate error amplification of high-hierarchy joints. Meanwhile, we design a hybrid loss function to realize dual supervision at the bone level and joint level, further improving the accuracy and structural rationality of pose estimation. 2. Related work We focus on diffusion-based 3D human pose estimation integrated with the core ideas of disentangled representation and hierarchical spatiotemporal modeling. Relevant research is mainly divided into two directions: 3D human pose estimation and diffusion models and their applications in pose estimation. Existing studies still have room for optimization in terms of human prior fusion, hierarchical information modeling, and diffusion model adaptation. The research progress and limitations of the two directions are reviewed. 2.1. 3D human pose estimation 3D Human Pose Estimation (3D HPE) is one of the core tasks in the field of computer vision, which strongly supports downstream applications such as virtual reality, human action recognition, and human-computer interaction. Its core goal is to regress the 3D spatial coordinates of human joints from monocular 2D pose sequences. Current mainstream methods can be categorized into one-stage methods and two-stage methods: One-stage methods directly extract features from raw RGB images and regress 3D poses. Typical works [,] use convolutional neural networks to model image feature volumes for end-to-end pose prediction, yet the estimation accuracy in the depth dimension is difficult to guarantee due to factors such as image background and human occlusion. Two-stage methods [,] first detect 2D keypoints in images using mature 2D pose estimation algorithms, and then complete 3D pose inference through 2D-to-3D lifting. This method has more advantages in accuracy and robustness as it fully leverages the mature technology of 2D pose estimation, and thus has become the current mainstream research paradigm. To alleviate the depth ambiguity problem in the 2D-to-3D mapping process, existing two-stage methods are mainly optimized from two perspectives: spatiotemporal information fusion and introduction of human pose priors. In terms of spatiotemporal information fusion, early studies used convolutional networks to mine temporal information to make up for the information loss in the depth dimension. Subsequent Transformer-based methods further fuse spatial and temporal context information and improve the accuracy of 3D pose estimation by capturing the spatiotemporal correlation of keypoints, with representative works including MixSTE [], PoseFormer [], and PoseFormerV2 []. For example, Zhang et al. [] proposed MixSTE (Mixed Spatio-Temporal Encoder), which consists of alternating temporal and spatial transformer blocks to separately model each joint’s temporal motion and inter-joint spatial correlation, and extends the output to the entire input video sequence for better coherence. Zheng et al. [] proposed a purely transformer-based model for 3D human pose estimation in videos (PoseFormer), which designs a spatial-temporal transformer structure to model intra-frame human joint relations and inter-frame temporal correlations without using any convolutional architectures. Although such methods exhibit strong capabilities in temporal modeling, they generally pay insufficient attention to the fine-grained hierarchical information of the human skeleton. The human skeleton presents a tree-like kinematic structure, and keypoints can be divided into different hierarchies according to kinematic depth. Existing methods ignore such hierarchical correlation, leading to the continuous accumulation of errors with the increase of hierarchies and ultimately affecting the overall estimation performance. In terms of introducing human pose priors, a class of disentangled methods has become an important research branch. These methods decompose 3D pose estimation into independent prediction tasks of bone length and bone direction, and then reconstruct 3D joint coordinates based on the forward kinematics of the human skeleton, which can explicitly incorporate human anatomical priors such as temporal consistency of bone length, joint angle limits, and human symmetry loss constraints, with representative works including Anatomy3D [], DKA [], and Virtual Bones []. For example, Chen et al. [] proposed a novel 3D human pose estimation model that decomposes the task into bone direction prediction and bone length prediction inspired by human skeleton anatomy. Xu et al. [] proposed a deep kinematics analysis (DKA) pipeline that systematically integrates 2D kinematics structure optimization, human-topology-based articulated motion decomposition, and a temporal refinement module to model both static and dynamic structures for monocular 3D pose estimation. Although disentangled methods can effectively fuse human priors, they have a key problem of hierarchical error amplification: since the 3D joint coordinates are derived layer by layer based on the bone parameters of parent and child joints, the prediction errors of bone length and direction will accumulate continuously in hierarchical propagation, ultimately resulting in their performance being significantly lower than that of current mainstream non-disentangled methods, which also becomes a core bottleneck for the practical application of disentangled methods. 2.2. Diffusion models As a class of generative models, diffusion models have achieved breakthrough results in image generation, super-resolution, semantic segmentation, multi-modal tasks and other fields by virtue of their strong probabilistic modeling and denoising capabilities. The core framework of the model includes a forward diffusion process and a reverse denoising process: the forward process gradually adds Gaussian noise to the original data until the data eventually degenerates into a random noise distribution; the reverse process recovers the original data from the noisy data by training a denoising model. Classic works such as DDPM (Denoising diffusion probabilistic models) [] and DDIM (Denoising Diffusion Implicit Models) [] have simplified and accelerated the training and sampling process of diffusion models, laying a foundation for their applications in various fields. In recent years, researchers have begun to explore the application of diffusion models in the 3D human pose estimation task, aiming to solve the uncertainty problem in 3D pose estimation by using their probabilistic modeling capabilities. Most existing diffusion-based 3D HPE methods [,] take the original 3D pose as the diffusion object, directly add noise to the 3D joint coordinates in the forward process, and then recover the real pose from the noisy 3D pose through the reverse denoising process. For example, Liu et al. [] proposed a novel Uncertainty-Guided Diffusion Model for 3D Human Pose Estimation (UDM-HPE) that uses uncertainty-aware probabilistic guidance to improve the denoising process. Some works also combine strategies such as multi-hypothesis aggregation, gradient field prior, and heatmap distribution initialization to improve performance, with representative works including D3DP [], GFPose [], and DiffPose []. For example, Shan et al. [] proposed a novel probabilistic 3D human pose estimation method named D3DP (Diffusion-based 3D Pose estimation) combined with JPMA (Joint-wise re-Projection-based Multi-hypothesis Aggregation), where D3DP generates multiple 3D pose hypotheses via a diffusion model conditioned on 2D keypoints, and JPMA aggregates these hypotheses joint-by-joint using reprojection errors to produce the final pose. Although such methods can use the advantages of diffusion models to improve the robustness of pose estimation, they have a key problem of insufficient fusion of explicit human priors: directly diffusing 3D joint coordinates without combining the disentangled representation of the human skeleton makes it difficult to explicitly incorporate human anatomical priors such as bone length and direction, resulting in the model being unable to fully utilize the inherent constraints of the human structure and limiting the further improvement of estimation accuracy. At the same time, existing diffusion-based 3D HPE methods still inherit the problems of traditional Transformer methods, i.e., insufficient exploration of hierarchical spatiotemporal information of the human skeleton. The design of the denoising model does not consider the hierarchical correlation of keypoints and cannot effectively alleviate the problem of hierarchical error accumulation. In addition, the error accumulation problem of disentangled methods and the insufficient prior fusion problem of diffusion models are independent of each other. 3. Methodology This section elaborates on the technical details of the proposed HADD, which integrates disentanglement strategies into diffusion models and leverages hierarchical spatio-temporal modeling to mitigate error accumulation. The overall pipeline of HADD is illustrated in , which consists of four core components: a 3D pose disentanglement strategy that decomposes joint coordinates into bone length and direction, a forward diffusion process with dimension-wise noise injection, a reverse diffusion process empowered by a Hierarchical Spatial and Temporal Denoising Module (HSTD), and a loss function that unifies disentanglement supervision and pose regression constraints. All mathematical formulations are rigorously derived and explained, with notations and definitions standardized for consistency with the state-of-the-art literature in diffusion models and 3D HPE. Ethics statements. This study only uses publicly available and strictly de-identified academic benchmark datasets (Human3.6M and MPI-INF-3DHP), and no new human subject data collection was conducted. Both datasets have passed the ethical review of their original publishing institutions, all subjects have signed informed consent forms, and the data do not contain any personally identifiable information. The use of the datasets in this study strictly follows the ethical requirements and usage agreements at the time of their release, and complies with the academic ethical norms in the field of computer vision. Therefore, no additional ethical review approval is required. 3.1. Preliminaries Definition 1 (Joint hierarchy division rule). We select 17 human core joint points as the pose representation units, based on the following rationale []: this joint set constitutes the minimal necessary set of core joints in line with human anatomy and kinematics, which can fully cover the key nodes of the motion chain of the trunk, limbs and head. In addition, the 17 joint points can be naturally decomposed into 16 bone segments, which is well adapted to the feature decomposition of bone length and bone direction in the disentangled diffusion strategy. We divide the 17 human joints into 6 hierarchies according to the depth from the root joint of the kinematic tree, with the following specific division, as shown in : The division strictly abides by the inherent tree-like kinematic structure of the human skeleton, taking the pelvic root node as the root and classifying the joints step by step from the proximal trunk to the terminal limbs according to their depth in the kinematic tree, which accurately matches the parent-child dependency among human joints. Definition 2 (Pinhole Camera Projection Model). We adopt the standard pinhole camera model for 3D-to-2D projection, which is fully compatible with the camera calibration parameters provided by the Human3.6M and MPI-INF-3DHP datasets. For the 3D joint coordinate Yn,j∈R3 (n∈[1,N] is the frame index, j∈[1,J] is the joint index), its 2D projection coordinate Yn,j′∈R2 on the image plane is jointly determined by the camera intrinsic and extrinsic parameters. The projection formula is as follows: where K∈R3×3 is the camera intrinsic matrix provided by the dataset (containing focal length and principal point coordinates), R∈R3×3 and g∈R3×1 are the camera extrinsic parameters (rotation and translation matrices). π(·) is a conversion function from homogeneous coordinates to inhomogeneous coordinates, defined as: For a complete 3D pose sequence Y∈RN×J×3 containing N frames and J joints, the projected 2D pose sequence is Y′∈RN×J×2, which is completely consistent in dimension with the input 2D pose sequence U∈RN×J×2, providing a unified dimensional basis for subsequent reprojection error calculation. We have established a unified key symbol definition system for the full text (in ). It should be mentioned that the explanations of temporary variables are provided at their first appearance in the text. 3.2. 3D pose disentanglement strategy The fundamental motivation of the disentanglement strategy is to decouple the high-dimensional and densely correlated 3D joint coordinate optimization problem into low-dimensional, physically interpretable subproblems of bone length and bone direction prediction. This decomposition not only simplifies gradient-based optimization but also naturally embeds human anatomical priors (e.g., temporal consistency of bone length, kinematic constraints of bone direction) into the diffusion model, addressing the limitation of vanilla diffusion-based 3D HPE methods that fail to learn explicit human pose priors due to direct noise injection on 3D joint coordinates. For a human skeletal structure with J joints, the 3D pose of a single frame is represented as y0∈RJ×3, where each element denotes the 3D coordinate of a joint in the Cartesian space. For a video sequence with N frames, the ground-truth 3D pose sequence is defined as Y0. Each bone is uniquely associated with a parent joint pi and a child joint ci. For the i-th bone, its ground-truth length L0i and unit direction vector D0i are mathematically formulated as: where ‖·‖2 denotes the Euclidean L2 norm, which quantifies the physical length of the bone between the parent and child joint. The Y0pi, Y0ci∈RN×J×3 represent the 3D coordinates of the parent and child joints of the i-th bone in the ground-truth pose. The unit direction vector is normalized to have a magnitude of 1, thus isolating the directional information of the bone from its length. For the entire video sequence, the disentangled ground-truth bone length sequence and bone direction sequence are denoted as L0 and D0, respectively. The disentanglement of 3D poses into L0 and D0 confers two critical properties for diffusion-based 3D HPE. Properties of the disentangled representation are as follows: (1) The original optimization space of N×J×3 is split into two independent spaces of N×B×1 (bone length) and N×B×3 (bone direction), reducing the number of trainable parameters and alleviating the problem of vanishing gradients in high-dimensional space; (2) Bone length is a time-invariant scalar for a given human subject (i.e., L0i remains constant across video frames), while bone direction is constrained by physiological joint angle limits. These priors can be explicitly enforced via loss functions, ensuring the physical plausibility of the predicted 3D pose. From a theoretical perspective, the rationality of confining disentanglement to the forward diffusion process is rooted in the inherent separability and heterogeneity of human skeletal motion features: (1) Human skeletal motion follows a tree-like hierarchical kinematic chain, where bone length (a time-invariant scalar) and bone direction (a spatiotemporally dynamic 3D vector) exhibit distinct statistical distributions. Bone length remains nearly constant across frames for a specific individual, while bone direction is constrained by physiological joint angle limits and exhibits motion-pattern-dependent variation. This inherent separability ensures that disentangled features can be independently modeled without information loss; (2) According to the principle of disentangled representation in information theory, decomposing the joint coordinate space into bone length and bone direction spaces minimizes the mutual information between features, reducing the model’s learning complexity from optimizing a high-dimensional joint distribution p(Y|U) to two low-dimensional marginal distributions p(L|U) and p(D|U) (U denotes input 2D pose sequences). This dimensionality reduction alleviates vanishing gradients in high-dimensional optimization and accelerates convergence. 3.3. Forward diffusion process The forward diffusion process of HADD follows the Markov chain formulation of Denoising Diffusion Probabilistic Models (DDPM) [], and we apply the dimension-wise Gaussian noise injection to the disentangled bone length L0 and bone direction D0 instead of the original 3D joint coordinates. This operation enables the diffusion model to learn the temporal dynamics of bone length and direction separately, preserving human anatomical priors during the noise addition process. For a generic data sample x0, the forward diffusion process gradually adds independent Gaussian noise to x0 over T steps, generating a noisy sample xt at step t (1≤t≤T). The process is defined as a Markov chain with a fixed noise schedule (the formula format consistent with the DDPM), where the conditional probability distribution of xt given x0 is: where N(·;μ,Σ) denotes the multivariate Gaussian distribution with mean μ and covariance matrix Σ; α ̄t is the cumulative product of αs from step 0 to step t; βs is the noise schedule coefficient35 that satisfies 0<β1<β2<⋯<βT<1; I is the identity matrix of appropriate dimension, ensuring noise is independently added to each dimension of the data. At step t, the noisy sample xt can be directly derived from x0 (bypassing intermediate steps) using the reparameterization trick: where ε is a standard Gaussian noise sample independent of x0. Applying to the disentangled bone length sequence L0 and bone direction sequence D0, the noisy bone length Lt∈RN×B×1 and noisy bone direction Dt∈RN×B×3 at step t are obtained as: where εl~N(0,I) and εd~N(0,I) are independent standard Gaussian noise samples for bone length and bone direction, respectively. The use of separate noise samples ensures that the noise dynamics of bone length (a scalar) and bone direction (a 3D vector) are modeled independently, avoiding cross-dimensional interference that would otherwise obscure anatomical priors. As t→T, α ̄t→0 and (1−α ̄t)→1, this means Lt and Dt converge to pure Gaussian noise. This property is consistent with DDPM and ensures the reverse diffusion process can be trained to recover the original disentangled features from random noise. 3.4. Reverse diffusion process The reverse diffusion process is the inverse of the forward process, aiming to recover the original disentangled bone length L0 and bone direction D0 from the noisy samples Lt and Dt (and ultimately from pure Gaussian noise at inference time). Unlike the forward process, which has a fixed closed-form distribution, the reverse process is a learned conditional distribution pθ, where θ denotes the trainable parameters of the model and U is the input 2D pose sequence (the conditional context for 3D HPE). The HSTD (Hierarchical Spatial and Temporal Denoising) module is the core of pθ, which models the hierarchical spatio-temporal correlations of human joints to denoise Lt and Dt effectively. Our Hierarchical Spatial and Temporal Denoising (HSTD) module, combined with the regression head, explicitly predicts the clean 3D joint coordinate sequence Y^0 as its only direct output. It does not predict the Gaussian noise term as in vanilla DDPM, nor does it directly predict the clean disentangled bone length L^0 and bone direction D^0. In the training stage, the reverse process takes the concatenated noisy features [Lt,Dt]∈RN×B×4 and the input 2D pose sequence U as inputs, and outputs the denoised 3D joint sequence Y^0 via the HSTD module and a regression head. Y^0 is the final direct output of the denoising network, while L^0 and D^0 are derived by secondary decomposition of Y^0 using . The conditional distribution of the reverse denoising process is formulated as: where Pθ(·) denotes the HSTD network with trainable parameters θ, and σt is the standard deviation of the additional Gaussian noise introduced in the DDIM reverse sampling procedure, with σt2 being its corresponding variance. This parameter controls the stochasticity of the sampling process and the variance of the predicted 3D pose distribution. t denotes the current diffusion timestep. Remark. Camera intrinsic and extrinsic parameters (K, R, g) are not conditions of the denoising network. They are only introduced in the multi-hypothesis selection stage at inference time to compute 2D reprojection errors for hypothesis ranking, and do not participate in the reverse diffusion denoising computation. The ground-truth 3D pose sequence is denoted as Y0, and the last dimension corresponds to the 3D Cartesian coordinates (x, y, z) of each joint. The denoised 3D pose sequence predicted by the model is denoted as Y^0. For both Y0 and Y^0, we perform disentanglement based on the human skeletal kinematic tree to obtain the ground-truth bone length L0, ground-truth bone direction D0 and predicted bone length L^0, predicted bone direction D^0, respectively. The disentanglement operation for the i-th bone (connecting parent joint pi and child joint ci) follows the Euclidean norm and unit vector normalization as defined: where Y^0pi, Y^0ci∈R3 are the corresponding predicted coordinates of the 3D coordinates of the parent and child joints of the i-th bone in the ground-truth pose. The disentangled bone length is a scalar (L0, L^0) and the bone direction is a 3D unit vector (D0, D^0). In the inference stage of the reverse diffusion process, to address the inherent depth ambiguity problem in monocular 2D-to-3D human pose lifting and further improve the accuracy and robustness of 3D pose prediction, our method draws on the multi-hypothesis sampling strategy proposed in D3DP [], and combines the fast-sampling advantage of DDIM [] to optimize the entire inference pipeline. Different from the fixed one-to-one denoising mode in the training stage, the inference stage adopts a multi-hypothesis iterative sampling mechanism for the disentangled bone length and direction features, which fully explores the multi-solution space of 2D-to-3D lifting and effectively avoids the prediction deviation caused by single random noise sampling. Specifically, we first sample H independent hypothesis samples of initial noisy bone length and initial noisy bone direction from the standard Gaussian distribution N(0,I) at the same time, where the number of hypotheses H is a tunable hyper-parameter that determines the coverage of the solution space. These H groups of initial noisy bone length and direction features are then input into the pre-trained HSTD for the first round of denoising processing. Through the hierarchical spatio-temporal feature modeling and denoising capability of the HSTD, we obtain the denoised bone length set L^0(h) and denoised bone direction set D^0(h) corresponding to the H hypotheses. Following the standard DDIM sampling paradigm [], we derive the timestep transition from the current step t to the next sampling step t* (t*