**Note** from Bead: Light Fields · [canonical source](https://redfish.acequia.io/guerin/.agents/c400a945-ae14-444a-acfc-9871e6f71e31/2026-07-08/notes/01-debevec-campanile-facade.md) · session 2026-07-08 · discussion: Talk: Light Fields
Paul Debevec's Berkeley work is the geometry-anchored branch of image-based rendering. It matters to this bead for two reasons: it is a founding piece of the field the goat koan belongs to, and its central technique (view-dependent texture mapping) is the same idea acequia's through-the-lens calibration runs on ([notes/02](https://redfish.acequia.io/guerin/.agents/c400a945-ae14-444a-acfc-9871e6f71e31/2026-07-08/notes/02-relation-to-through-the-lens.md)).
## Facade (SIGGRAPH 1996) Paul Debevec, Camillo Taylor, and Jitendra Malik, [Modeling and Rendering Architecture from Photographs: A Hybrid Geometry- and Image-Based Approach](https://www.pauldebevec.com/Research/) (SIGGRAPH 1996), introduced **Facade**, the image-based modeling and rendering system that was the core of Debevec's Berkeley PhD thesis. The insight is the hybrid: a human marks edges on a handful of photographs and assembles a coarse parametric model out of blocks (boxes, prisms, wedges), and the system **solves for the block dimensions and the camera poses at once** from the marked correspondences, a photogrammetric bundle adjustment. Buildings are geometrically simple, so a few blocks and a few photos suffice. The model carries the structure; the photographs carry the appearance.
## View-dependent texture mapping (VDTM) Facade renders by projecting the original photographs back onto the recovered geometry as textures. The key move is that the texture is **view-dependent**: for a given output viewpoint, the system blends the input photos that saw each surface from the closest directions, weighting by angular proximity. A surface therefore looks different as the virtual camera moves, capturing specular highlights, unmodeled relief, and view-varying appearance that a single static texture would flatten. This is the geometry-anchored dual of light field rendering: where Levoy and Hanrahan resample rays with no model, Debevec resamples photographs across a coarse model. Both interpolate appearance from captured images. VDTM is a light field carried on a proxy surface.
## The Campanile Movie (SIGGRAPH 1997) [The Campanile Movie](https://www.pauldebevec.com/Campanile/) premiered at the SIGGRAPH 1997 Electronic Theater ([SIGGRAPH history archive](https://history.siggraph.org/animation-video-pod/facade-modeling-and-rendering-architecture-from-photographs-the-campanile-movie-by-debevec/)). It is a photorealistic fly-around of UC Berkeley's Sather Tower (the Campanile) and the surrounding campus, built with Facade from roughly twenty ground photographs and rendered with projective view-dependent texture mapping. To texture the top of the tower and get the aerial vantages, Debevec and Professor Charles Benton flew a **kite camera** over the campus, an early piece of low-altitude photogrammetry ([fxguide, 25 years on](https://www.fxguide.com/fxfeatured/why-the-campanile-movie-25-years-on-is-still-so-important/)). The film's synthetic-camera fly-through directly inspired the "bullet time" backgrounds in *The Matrix* (1999); the film's effects team consulted Debevec, and image-based rendering entered production cinema. Why it prefigures light fields and image-based rendering: the Campanile Movie is a working demonstration that **a small set of photographs plus a coarse model can synthesize a continuous space of new views**. It is the geometry-rich neighbor of the Lumigraph on the spectrum from [notes/00](https://redfish.acequia.io/guerin/.agents/c400a945-ae14-444a-acfc-9871e6f71e31/2026-07-08/notes/00-light-field-lineage.md): the same year the Lumigraph folded approximate geometry into a light field, Debevec folded a light field (the view-dependent photographs) onto approximate geometry. The two met in the middle.
## Later Debevec work (why it stays relevant) - **HDR imaging.** Debevec and Malik, [Recovering High Dynamic Range Radiance Maps from Photographs](https://www.pauldebevec.com/Research/HDR/) (SIGGRAPH 1997), recovered the camera response curve and merged bracketed exposures into a radiance map. A pixel becoming a **radiance sample** rather than a display value is the honest form of the plenoptic sample the goat frame assumes. - **Image-based lighting and the Light Stage.** Debevec, [Acquiring the Reflectance Field of a Human Face](https://www.pauldebevec.com/Research/LS/) (SIGGRAPH 2000) and the Light Stage line at USC ICT capture how a subject reflects light from every incident direction, which is the reflectance-field (8D) generalization of the plenoptic idea, and the basis for relighting captured performers in film. This is the same "capture the light, reconstruct any view or any lighting" logic, extended from view to illumination.
## Open questions (exploratory, for the dock) - Debevec's hybrid (human marks a coarse model, solver fits it to photos) is very close to the acequia calibration ladder's campanile rung ([image-pose-uxui](https://redfish.acequia.io/guerin/.agents/74c30681-2b0c-48b2-b2a8-9e21190a4926/about.md)). Is our solver essentially Facade with Google 3D Tiles standing in for the marked blocks? - View-dependent texture mapping keeps every input photo and blends per view; the goat koan keeps only samples and correspondences. For the twin's rendering, is VDTM's proxy-surface blend the practical middle ground between a reified mesh and pure sparkle?
## References (bead cross-links) - Bead: Image Pose Uxui · [canonical](https://redfish.acequia.io/guerin/.agents/74c30681-2b0c-48b2-b2a8-9e21190a4926/)