Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics
Authors: Changmin Lee, Jihyun Lee, Tae-Kyun Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we demonstrate that MPMAvatar significantly outperforms the existing state-of-the-art physics-based avatar in terms of (1) dynamics modeling accuracy, (2) rendering accuracy, and (3) robustness and efficiency. Additionally, we present a novel application in which our avatar generalizes to unseen interactions in a zeroshot manner which was not achievable with previous learning-based methods due to their limited simulation generalizability. Our project page is at: https: //KAISTChangmin.github.io/MPMAvatar/. |
| Researcher Affiliation | Academia | Changmin Lee Jihyun Lee Tae-Kyun Kim KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Collision Handling |
| Open Source Code | No | Our code will be also publicly available to allow full reproducibility. |
| Open Datasets | Yes | We perform our main evaluations on (1) Actors HQ [14]. In particular, we select four subjects used in [78]: two characters in loose dresses and two characters in two-piece outfits. ... we additionally include four sequences from (2) 4D-DRESS [66] dataset, to perform more extensive comparisons. |
| Dataset Splits | Yes | For training each subject, we use 24 frames with large cloth dynamics for physical parameter learning and 200 frames for appearance learning. For testing, we use 200 unseen frames per subject. Whereas the existing work [70] only uses Actors HQ for evaluation, we additionally include four sequences from (2) 4D-DRESS [66] dataset, to perform more extensive comparisons. We use two subjects in tops and skirts and two in tops and tight jeans. For training, we use 11 frames for physical parameter learning and 100 frames for appearance learning, while testing was carried out on 100 unseen frames. |
| Hardware Specification | Yes | Simulation time and computing resource. As noted in the main paper (Sec. 5.2), our simulation runs at approximately 1.1 seconds per frame on a single NVIDIA GeForce RTX 4090. |
| Software Dependencies | No | As the official implementation of the MPM solver for this constitutive model is not publicly available, we re-implemented it using PyTorch [50] and Warp [40], and plan to release our code to facilitate future research. |
| Experiment Setup | Yes | Our MPM simulation uses a time step of t = 0.04 with N = 400 substeps and a grid resolution of 200. We optimize the physical parameters over 200 iterations using the Adam optimizer. For finite-difference gradient estimation, the perturbation sizes are set to Δρ = 0.05, ΔE = 5, and Δα = 0.005. The corresponding learning rates are 0.01 for ρ, 0.3 for E, and 0.01 for α. All parameters are initialized as ρ = 1.0, E = 100, and α = 1.0 for physical parameter learning, while ν, γ, and κ are fixed at their default values of 0.3, 500, and 500, respectively. |