Diffusion Path Alignment for Long-Range Motion Generation and Domain Transitions
Authors: Haichao Wang, Alexander Okupnik, Yuxing Han, Gene Wen, Johannes Schneider, Kyriakos Flouris
Organizations: Shenzhen International Graduate School, Tsinghua University, China · Institute of Information Systems, Universität Liechtenstein, Liechtenstein · Department of Computer Science, New York University, USA · MRC Biostatistics Unit, University of Cambridge, United Kingdom
Long-range human movement generation remains a central challenge in computer vision and graphics. Generating coherent transitions across semantically distinct motion domains remains largely unexplored. This capability is particularly important for applications such as dance choreography, where movements must fluidly transition across diverse stylistic and semantic motifs. We propose a simple and effective inference-time optimization framework inspired by diffusion-based stochastic optimal control. Specifically, a control-energy objective that explicitly regularizes the transition trajectories of a pretrained diffusion model. We show that optimizing this objective at inference time yields transitions with fidelity and temporal coherence. This is the first work to provide a general framework for controlled long-range human motion generation with explicit transition modeling.
Figures & tables
Figure 2: Illustration of the guided denoising with additional control energy objective. Black line illustrations trajectory in latent space at t=T , red represents unconditional denoising trajectory with ϵθ(xt,t;∅) and the blue line with mixed guided denoising ϵθ(xt,t;ω) . Ec is the control energy to be minimized for the guidance mechanisms to be close to the unguided trajectory.
FIDk↓
FIDm↓
Divk↑
Divm↑
Ground Truth
8.25
11.47
14.01
21.97
Linear
174.32
181.85
6.13
20.48
Sigmoid
171.85
178.83
6.08
20.91
Sine
148.48
148.47
7.04
20.80
M-DPA (Ours)
128.01
137.45
7.06
18.88
Table 1: Comparison of baselines and our method. Bold indicates the best result.
Accel Mean
Accel Var
Jerk Mean
Jerk Var
Ground Truth
4.526
101.687
138.975
1.707e5
Sine
3.733
73.151
111.084
1.076e5
M-DPA
3.958
91.427
117.488
1.449e5
Table 2: Motion dynamics statistics across different settings.
Figure 3: Subplots show generations of linear (first row), Sine (second row) and M-DPA (last row) generated movement samples. The dash lined bounding box marks the transition phase. While the heuristic methods like linear and sine interpolation introduces additional folding of the charakter, the M-DPA shows more coherent transitions.
Figure 4: Subplots show the results of the ω optimization along different denoising steps for different class transitions. Both inbetween segments 2 and 3 show a similar pattern. ω peaks at denoising step 7 and then monotonically decreasing, consistently for all the class transitions 0→1 , 0→5 , 0→9 .
Figure 5: Evolution of control energy for both the Sine interpolation baseline and M-DPA.