The ap­pear­ance of com­plex, thick ma­ter­i­als like tex­tiles is de­term­ined by their 3D struc­ture, and they are in­com­pletely de­scribed by sur­face re­flec­tion mod­els alone. While volume scat­ter­ing can pro­duce highly real­ist­ic im­ages of such ma­ter­i­als, cre­at­ing the re­quired volume dens­ity mod­els is dif­fi­cult. Pro­ced­ur­al ap­proaches re­quire sig­ni­fic­ant pro­gram­mer ef­fort and in­tu­ition to design spe­cialpur­pose al­gorithms for each ma­ter­i­al. Fur­ther, the res­ult­ing mod­els lack the visu­al com­plex­ity of real ma­ter­i­als with their nat­ur­ally-arising ir­reg­u­lar­it­ies.

This pa­per pro­poses a new ap­proach to ac­quir­ing volume mod­els, based on dens­ity data from X-ray com­puted tomo­graphy (CT) scans and ap­pear­ance data from pho­to­graphs un­der un­con­trolled il­lu­min­a­tion. To mod­el a ma­ter­i­al, a CT scan is made, res­ult­ing in a scal­ar dens­ity volume. This 3D data is pro­cessed to ex­tract ori­ent­a­tion in­form­a­tion and re­move noise. The res­ult­ing dens­ity and ori­ent­a­tion fields are used in an ap­pear­ance match­ing pro­ced­ure to define scat­ter­ing prop­er­ties in the volume that, when rendered, pro­duce im­ages with tex­ture stat­ist­ics that match the pho­to­graphs. As our res­ults show, this ap­proach can eas­ily pro­duce volume ap­pear­ance mod­els with ex­treme de­tail, and at lar­ger scales the dis­tinct­ive tex­tures and high­lights of a range of very dif­fer­ent fab­rics like sat­in and vel­vet emerge auto­mat­ic­ally—all based simply on hav­ing ac­cur­ate meso­scale geo­metry.