Stretch a cloth and its yarn gaps appear on their own: DTM compiles dynamic microgeometry into a neural texture, skipping per-frame reconstruction
Synopsis
Deformable Texture Maps (DTM) compile deformation-driven yarn-scale microgeometry offline into one continuous neural texture, so at runtime a texture coordinate plus stretch and shear parameters yield opacity, normal, depth and material-identity maps directly on coarse cloth meshes. Earlier paths either kept yarn geometry in the runtime representation, coupling reconstruction to rendering, or represented only static fabric appearance without the spacing and occlusion changes that deformation produces. Across five woven and knitted structures, a width-128 model reconstructs between ASTC 4x4 and ASTC 6x6 texture compression while cutting storage by 42x to 94x.
Interpretation
The detailed look of deformable surfaces such as cloth and skin comes from tiny geometry that rearranges as the surface deforms; DTM moves that detail out of runtime geometry reconstruction and into a texture queried by deformation. Earlier yarn-level methods kept fine-scale geometry in the runtime representation, coupling reconstruction and rendering to that microgeometry, while surface-based fabric appearance models represented static states without deformation-driven changes in spacing and occlusion. On five woven and knitted structures, rib, basket, stockinette, honeycomb and satin, the width-128 model's reconstruction quality falls between ASTC 4x4 and ASTC 6x6 compression of the whole training dataset, with 42x to 94x less storage.
The build sweeps deformation space, pulls each deformed yarn configuration back onto one rest-space texture footprint and renders it into maps, then a coordinate network compresses those samples into a continuous function. The coordinate network concatenates texture coordinate and deformation state, uses a fixed Fourier encoding with sinusoidal-activation SIREN to jointly predict eight channels across four maps, and trains on the equally weighted mean squared error of the four maps. Training samples come from physical relaxation of a tileable yarn patch over a stretch and shear grid, 512 deformation states in all; at runtime the network is queried on coarse garment meshes by interpolating per-vertex deformation parameters, and predictions are cached in texture space and refreshed only where deformation crosses a threshold.
Perspective
The demonstrated scope is yarn-based textiles: five woven and knitted structures, a bounded deformation domain described by local in-plane stretch and shear, and real-time rendering and editing on coarse garment meshes.
Each DTM trains for a single microstructure, and what is demonstrated is interpolation between sampled deformation states, not extrapolation beyond the domain or transfer across yarn structures; the deformation descriptor covers in-plane stretch and shear but not bending or localized compression; update frequency and row interleaving can delay changes in the predicted yarn fields, and judging that approximation requires comparison with full per-frame texture-space updates at the same resolution.
