UNIC: Neural Garment Deformation Field for Real-time Clothed Character Animation

How to read this

Category

real time neural garment deformation paper

Contributions

• Proposes an instance specific neural deformation field, an MLP mapping arbitrary 3D points to deformation offsets, instead of a generalising graph neural network, trading cross garment generalisation for high fidelity on one character and garment across new motions
• Avoids explicit topology handling by operating in continuous 3D space rather than on mesh vertices or graph edges, letting it animate complex garments (dresses, layered clothing) that graph based methods fail on because they are bound to the topologies seen in training
• Encodes the character's pose history into a discrete, categorical motion latent space before querying the deformation field, intended to regularise generalisation to unseen motions and prevent drift over long autoregressive sequences
• Adds a lightweight intersection handling module to resolve garment to body penetration
• Reports real time inference on a single RTX 3090 with quality comparable to GPU accelerated professional cloth simulation software, evaluated on SMPL characters wearing garments from t shirts to complex dresses

Context

UNIC positions itself against both physics based cloth simulators, which are too slow for real time use, and prior graph neural network based learned cloth deformation, which the authors argue is fundamentally limited by the fixed topologies it was trained on. Framing garment deformation as a continuous neural field rather than a per vertex graph prediction follows the broader neural fields trend, here applied to instance specific cloth animation rather than scene reconstruction.

Correctness

The claims rest on comparison against unnamed professional cloth simulation software as the quality and speed baseline, and the instance specific design is an explicit and stated tradeoff: the model does not generalise to new garments or characters without retraining, which limits it to per asset production pipelines rather than a general purpose tool. The drift prevention claim depends on the categorical motion latent trick, and how robust that is over very long sequences is asserted in the introduction but not detailed there, it would need checking in the experiments section.

Clarity

The problem framing, graph neural network topology limits versus neural field flexibility, is clear and accessible to a technical artist. The deeper architecture, loss functions and training details would need the method section to assess fully.

How to read it

First pass, read the abstract and Figure 2, the quality and efficiency comparison to the gold standard, to see exactly what tradeoff is being offered. Second pass, read the introduction's explanation of the instance specific paradigm shift and the categorical motion latent space, since that is the core technical bet of the paper. Third pass, read the method and experiments sections for the network architecture, training data, and the actual comparison numbers against professional software and prior graph neural network baselines.