Digital Humans, Avatars & Virtual Try-On
Reducing Downstream Dependency of Garment/Avatar Generation on Upstream 2D and 3D Human-Parsing Predictions
Generated automatically from the limitations stated in 8 papers (ICCV, CVPR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Garment and avatar pipelines that degrade gracefully on exaggerated poses, unusual garments, and image types the off-the-shelf estimators were never trained for, without requiring bespoke compensating steps.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Error-propagation audit: take one existing garment/avatar pipeline, inject controlled perturbations (keypoint jitter, mask erosion, SMPL mis-fit) at each preprocessing stage, and measure output degradation (garment IoU, leaking pixels, per-vertex error) to produce the first quantitative sensitivity map of where pipeline failures originate.
- 2
Confidence-aware consumption of upstream predictions: augment parsing/pose outputs with per-pixel/per-joint uncertainty and train the downstream garment model to weigh or fallback on uncertain regions, measuring whether leaking and shape errors drop on perturbed and out-of-distribution inputs.
- 3
Joint fine-tuning: co-train the parser and the garment generator so preprocessing adapts to downstream needs, evaluated on poses and garment types underrepresented in the original training sets.
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
The perturbation study may reveal errors are irreducibly idiosyncratic per pipeline — each fails in ways too different to share a fix — leaving only a diagnostic benchmark rather than a general remedy.
Sub-problems
Robust Virtual Try-On and Clothed Human Modeling Under Upstream Preprocessing Noise
Current virtual try-on and clothed avatar pipelines rely on a brittle multi-stage preprocessing cascade comprising 2D garment parsing, SAM segmentation, and SMPL/SMPL-X body fitting. Upstream estimation errors directly propagate into the downstream generation stage, causing severe visual artifacts such as garment leakage, unnatural boundary tearing, and shape distortion. Because existing pipelines assume clean segmentation masks and accurate parametric fits, they fail completely on unconstrained real-world imagery featuring loose clothing, complex poses, or partial occlusions.
Evidence
Each paper's own statement of the limitation, verbatim.
- Pose-Star: Anatomy-Aware Editing for Open-World Fashion ImagesICCV 2025
Requires pose keypoints from OpenPose, so performance depends on pose-estimation quality and the method is specific to human/fashion imagery (body-structure decomposition of instructions)
- OmniVTON: Training-Free Universal Virtual Try-OnICCV 2025
Relies on OpenPose keypoint and TAPPS segmentation predictions; failure of these off-the-shelf estimators (e.g., unparseable garment-only images) requires a compensating pseudo-person generation step that itself may fail.
- NGD: Neural Gradient Based Deformation for Monocular Garment ReconstructionICCV 2025
Requires heavy pre-trained priors (4DHumans, Sapiens, BCNet) for poses, pseudo-ground-truth maps, and initial garment template, so performance depends on their accuracy
- DAGSM: Disentangled Avatar Generation with GS-enhanced MeshCVPR 2025
Relies on SAM segmentation quality and SMPL-X body prior; garment initialization from body regions may constrain garment shapes, and only limited garment/body types were evaluated (10 avatars)
- TryOnDiffusion: A Tale of Two UNetsCVPR 2023
Garment leaking artifacts occur when preprocessing (parsing/pose estimation) errors produce bad segmentation maps or garment leakage into the clothing-agnostic RGB image
- REC-MV: REconstructing 3D Dynamic Cloth From Monocular VideosCVPR 2023
Relies on preprocessing: SMPL fitting via Videoavatar, garment parsing masks, and 2D visible curves from mask boundaries; errors in these propagate
- Dynamic Point FieldsICCV 2023
Does not model dependency of clothing shape on body pose or motion acceleration (no pose-space generalization; reposing uses nearest training scan for best results)
- Controllable Person Image Synthesis with Pose-Constrained Latent DiffusionICCV 2023
Fails on poses exaggerated relative to the training set, underrepresented garments, noisy DensePose annotations, and garment changes between source and target
Nearest existing work
- CLOTH3D: Clothed 3D HumansECCV 2020
- Clothe and PoseCVPR 2026
- 4D-DRESS: A 4D Dataset of Real-World Human Clothing With Semantic AnnotationsCVPR 2024
- SCANimate: Weakly Supervised Learning of Skinned Clothed Avatar NetworksCVPR 2021
- Time-Efficient and Identity-Consistent Virtual Try-On Using A Variant of Altered Diffusion ModelsECCV 2024
- Virtual Try-On with Pose-Garment Keypoints Guided InpaintingICCV 2023
- Cloth2Body: Generating 3D Human Body Mesh from 2D ClothingICCV 2023
- OmniVTON: Training-Free Universal Virtual Try-OnICCV 2025
- SwapNet: Garment Transfer in Single View ImagesECCV 2018
- A Generative Model of People in ClothingICCV 2017
- DiffAvatar: Simulation-Ready Garment Optimization with Differentiable SimulationCVPR 2024
- Size Does Matter: Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on NetworkICCV 2023
- Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single ImagesECCV 2020
- Garment4D: Garment Reconstruction from Point Cloud SequencesNeurIPS 2021
- Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-OnCVPR 2021
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
Joint Pose Refinement and Error-Tolerant Conditioning for Downstream Human Avatar Synthesis
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.
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.