Digital Humans, Avatars & Virtual Try-On
Joint Pose Refinement and Error-Tolerant Conditioning for Downstream Human Avatar Synthesis
Generated automatically from the limitations stated in 8 papers (CVPR, ICCV, ECCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current avatar creation and virtual try-on pipelines treat SMPL/SMPL-X estimations, camera calibrations, and segmentation masks as fixed, ground-truth preconditions. When off-the-shelf estimators inevitably produce misaligned or out-of-distribution body parameters on in-the-wild images, these errors cascade into severe geometric distortion, stitching artifacts, and generation failure. Because existing downstream pipelines do not jointly optimize human body parameters or account for parameter uncertainty, high-fidelity human generation remains restricted to controlled studio captures with pristine initializations.
Why it matters
Enables photorealistic digital human avatars and virtual try-on to run reliably on unconstrained monocular images and videos without requiring manual mask correction or calibrated multi-camera rigs.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Joint photometric-geometric refinement: Formulate avatar reconstruction to backpropagate rendering and perceptual losses into SMPL-X pose and camera parameters during downstream fitting rather than freezing upstream estimates, measuring surface reconstruction error (Chamfer distance, PVE) and image alignment on in-the-wild benchmarks.
- 2
Uncertainty-aware conditioning for generative try-on/avatar models: Inject probabilistic pose distributions or dense vertex correspondence maps instead of rigid SMPL meshes into generation networks, measuring synthesized image fidelity (FID, LPIPS) across controlled levels of synthetic and real pose estimation noise.
- 3
Iterative test-time pose adjustment via featuremetric alignment: Implement feature-space gradient descent on SMPL parameters guided by self-supervised neural rendering discrepancies, evaluating keypoint reprojection accuracy and artifact reduction across diverse lighting and occlusion conditions.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Foundation human mesh recovery models might improve to near-perfect pixel alignment on arbitrary in-the-wild inputs, dissolving the upstream error problem at the estimator stage before downstream refinement methods mature.
Evidence
Each paper's own statement of the limitation, verbatim.
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit RemeshingCVPR 2025
SMPL-X pose estimation errors cascade into generation and reconstruction, producing inaccurate poses and stitching artifacts
- Disco4D: Disentangled 4D Human Generation and Animation from a Single ImageCVPR 2025
Depends on robust pixel-aligned SMPL-X estimation, which is unsolved; fails when SMPL-X fitting or visual hull initialization is poor
- ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the WildECCV 2024
Relies on reasonable SMPL pose estimates and SAM segmentation masks as inputs; SAM masks occasionally require manual correction for sharp boundaries
- GauHuman: Articulated Gaussian Splatting from Monocular Human VideosCVPR 2024
Requires calibrated cameras, human masks, and pre-estimated SMPL shape/pose parameters as inputs
- SHERF: Generalizable Human NeRF from a Single ImageICCV 2023
Relies on assumed inputs: calibrated camera parameters, human masks, and ground-truth SMPL pose/shape parameters for the input image
- Neural Reconstruction of Relightable Human Model from Monocular VideoICCV 2023
Requires relatively precise initial SMPL pose parameters; inaccurate poses cause imprecise reconstruction and unrealistic decomposed parameters, especially under continuous illumination changes.
- Learning Neural Volumetric Representations of Dynamic Humans in MinutesCVPR 2023
Relies on accurate pre-computed SMPL parameters, which are difficult to obtain in in-the-wild settings; pose is not jointly optimized
- EVA3D: Compositional 3D Human Generation from 2D Image CollectionsICLR 2023
SMPL parameters estimated from 2D images are inaccurate, creating a distribution shift from real poses that may compromise generation
Nearest existing work
- Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human AvatarsICCV 2025
- Motion-Aware Animatable Gaussian Avatars DeblurringCVPR 2026
- Bringing Your Portrait to 3D PresenceCVPR 2026
- High-Fidelity Human Avatars From a Single RGB CameraCVPR 2022
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal PriorCVPR 2025
- GASP: Gaussian Avatars with Synthetic PriorsCVPR 2025
- GAS: Generative Avatar Synthesis from a Single ImageICCV 2025
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D GaussiansCVPR 2024
- DreamWaltz: Make a Scene with Complex 3D Animatable AvatarsNeurIPS 2023
- MoGA: 3D Generative Avatar Prior for Monocular Gaussian Avatar ReconstructionICCV 2025
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian ReconstructionCVPR 2025
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic DataCVPR 2024
- LHM: Large Animatable Human Reconstruction Model for Single Image to 3D in SecondsICCV 2025
- SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic PriorsCVPR 2025
- HUGS: Human Gaussian SplatsCVPR 2024
Related open problems
Digital Humans, Avatars & Virtual Try-On
Overcoming Parametric 3DMM Expressiveness and Mouth Interior Bottlenecks in Monocular Facial Avatars
Neural head avatar pipelines universally rely on parametric mesh models like FLAME as geometric proxies, inheriting their topological limitations. Because FLAME lacks internal oral anatomy (teeth and tongue) and fine-grained geometric detail, avatars suffer from severe projection artifacts such as teeth textures baked directly onto lip surfaces and distorted speech articulation. Furthermore, low-dimensional linear expression spaces cap the capture of dynamic micro-expressions, dynamic wrinkles, and non-parametric regions like hair.
Digital Humans, Avatars & Virtual Try-On
Reducing Downstream Dependency of Garment/Avatar Generation on Upstream 2D and 3D Human-Parsing Predictions
Every pipeline in this space chains off-the-shelf estimators — pose keypoint detectors, DensePose, SMPL/SMPL-X fitting, segmentation/parsing models, SAM — and treats their outputs as ground truth. When these upstream predictions are noisy, fail on exaggerated poses, unusual garments, or garment-only imagery, the downstream model inherits or amplifies the errors (garment leaking, misinitialized garment geometry, failed reposing). No one measures or controls how error propagates through this preprocessing chain, so each new pipeline is brittle in its own undocumented way.
Digital Humans, Avatars & Virtual Try-On
Benchmarking and Improving 3D Human Body Model Robustness Across Underrepresented Demographics and Non-Standard Morphologies
Existing 3D statistical body models and avatar pipelines are predominantly trained on narrow demographic subsets (such as adult European CAESAR scans) and evaluated solely on standard normative body templates. Consequently, downstream avatars and virtual try-on methods systematically produce distorted geometry, unnatural surface artifacts, and inaccurate anthropometric measurements on children, adolescents, diverse global populations, individuals with disabilities, and non-standard body proportions. Because existing benchmarks do not measure generalization across diverse body shapes, failure modes on these populations remain unquantified and unaddressed.