A new artificial-intelligence system could make it dramatically easier to build editable digital humans from ordinary video, allowing virtual characters to change clothes, retain realistic body motion and be assembled without retraining the underlying model. The framework, called MAC-RAD, separates a person captured in monocular RGB footage into reusable digital assets—such as skin, body geometry and garments—before recombining those components into a single animatable avatar. The work, published in Multimedia Tools and Applications, targets a persistent problem in computer vision: most systems can reconstruct a convincing person, or generate clothing variations, but struggle to treat the body and garments as independent three-dimensional objects that can be rearranged reliably.
The distinction matters because an avatar reconstructed as one indivisible surface is difficult to edit. A jacket, shirt or pair of trousers may be visually represented in the pixels, yet the system may not know where one garment ends and another begins, how far each layer should sit from the body, or how clothing should deform during movement. MAC-RAD is designed around modularity instead. Its first stage extracts clothing and body elements into a canonical three-dimensional space, a standardized coordinate system in which assets can be stored and reused. Its second stage composes those assets into a unified mesh that can be animated. In principle, the same extracted body could wear several combinations of garments, while the expensive reconstruction step would not need to be repeated for every new outfit.
The researchers, Cheng-Yen Yu, Po-Chen Chen and Li-Chen Fu of National Taiwan University, build the system around two algorithmic strategies. The first is a vertex-based confidence voting method intended to overcome the instability of frame-by-frame image segmentation. Modern foundation models can identify objects and regions in individual video frames, but their predictions may flicker as a person moves, lighting changes or a garment becomes partially occluded. A shirt might be assigned slightly different boundaries from one frame to the next, producing fragmented or inconsistent geometry when those masks are projected into three dimensions. MAC-RAD instead treats the segmentation results as evidence accumulated across the video. Each mesh vertex receives information from multiple two-dimensional masks, and the system combines those observations to estimate whether the vertex belongs to a particular clothing item.
This voting process is designed to create coherent, textured garment meshes rather than a collection of disconnected frame-specific guesses. A mesh is a surface made from vertices and faces, and its structure provides the geometric scaffold needed for animation and rendering. Texture information can then be associated with that surface so that the reconstructed clothing retains visual details from the source video. The canonical representation also helps stabilize assets across time: rather than reconstructing a new shirt independently in every frame, the algorithm seeks one consistent garment whose appearance and shape explain observations throughout the sequence. That temporal aggregation is especially important for sleeves, hems and other boundaries that can be difficult to identify from a single viewpoint.
The second major challenge appears when separate clothing assets are put back together. Real clothing is layered: skin lies beneath a shirt, a jacket may cover the shirt, and folds or loose fabric can extend away from the body. If independently reconstructed meshes are simply placed in the same space, they can interpenetrate, with one surface cutting through another. Such errors are immediately visible in a rendered avatar, particularly around collars, cuffs, waistlines and overlapping jackets. To address this, MAC-RAD uses a signed distance field, or SDF, to guide mesh layering. An SDF assigns each point in space a value describing its distance from a surface, with the sign indicating which side of that surface the point occupies. The resulting field acts as a potential landscape around the canonical body and helps the system enforce a plausible ordering between body and garments.
The SDF-guided method does not amount to a complete physical simulation of cloth, and the source article does not claim to reproduce every wrinkle or collision caused by real-world fabric. Instead, it provides geometric constraints for composition. By using the body as a spatial reference and assigning clothing layers an appropriate relationship to that reference, the system can reduce the visual conflicts produced by naive mesh assembly. The researchers report that this geometry-aware layering improves visual coherence compared with monolithic baselines, systems that handle the person as a single combined representation rather than explicitly separating and recombining the components. The improvement is significant for applications in which users need to alter one part of an avatar while preserving the rest.
MAC-RAD divides its workload into offline and online phases. Asset extraction from the original RGB video is a one-time computational cost. Once the skin and garments have been disentangled and stored in canonical form, the composition stage is intended to be fast, supporting rapid or real-time assembly without retraining the model. This design addresses a practical bottleneck in generative avatar systems. A model that must be optimized again each time a user changes a jacket or swaps an outfit may produce high-quality results but remain too slow or expensive for interactive use. A reusable asset library, by contrast, could allow clothing combinations to be generated through direct geometric operations and rendering, provided the assets are compatible with the target body and animation system.
The framework draws on several strands of recent digital-human research, including monocular human reconstruction, animatable layered assets, articulated body models, Gaussian-based representations and neural rendering. Monocular reconstruction is inherently underconstrained: a single camera records only two-dimensional projections, leaving hidden surfaces and true depth ambiguous. Systems must therefore infer three-dimensional shape from learned priors about human bodies, poses and clothing. MAC-RAD uses semantic understanding from foundation models while adding explicit geometric processing for disentanglement and composition. That combination reflects a broader shift in computer vision from merely generating plausible images toward constructing structured representations that can be edited, animated and reused.
The researchers evaluated the approach using the People Snapshot dataset, which was used for evaluation rather than as a foundation on which the method or its main conclusions depend. The article describes experiments showing that the proposed layering strategy produces more coherent results than monolithic alternatives, but the available source does not provide a numerical performance table or a single headline accuracy score. The work should therefore be understood as a systems contribution rather than evidence that arbitrary clothing can already be reconstructed perfectly from any casual video. Monocular footage can contain severe occlusion, motion blur, unusual poses and garments whose appearance changes substantially with folds. These conditions remain difficult for any method that must infer hidden geometry from limited visual evidence.
The potential applications nevertheless extend well beyond virtual fashion. Modular avatars could support virtual and augmented reality, interactive games, online social spaces, digital clothing visualization and personalized telepresence. A user might eventually select garments by name, image or text, then see them assembled on an avatar whose body, face and motion remain unchanged. The authors point to CLIP-guided retrieval as a possible next step: a language-and-image model could connect text prompts with stored assets, allowing commands such as “add a dark jacket” or “replace the trousers with athletic pants” to trigger retrieval and composition. That capability would add semantic control on top of the geometric system, although it remains a proposed extension rather than a demonstrated feature of the current work.
MAC-RAD’s most important idea is that a digital person need not be built as a single permanent object. By treating skin and garments as separable, structured assets, the system aims to make avatar creation more like assembling components than regenerating an entire human from scratch. The approach could help bridge the gap between photorealistic reconstruction and practical editing, especially where speed and repeated customization matter. Yet its success will depend on how well the extracted meshes preserve appearance, how robustly they handle hidden regions and how convincingly the resulting layers behave during animation. For now, the study offers a geometry-aware route toward reusable virtual humans—and a glimpse of avatars that can be redesigned as easily as changing clothes.
