JHU and Adobe open-source system fusing inconsistent multi-view renderings into one seamless texture

ssh4net · x · 2026-10-29

Researchers from Johns Hopkins University and Adobe Research/Firefly Foundry (Hongyi Liu, Yifan Wang, Vladimir G. Kim, Misha Kazhdan) released a paper on texture consolidation of generated multi-view renderings, with BSD-3 licensed open-source code.

The system tackles a common problem: generative multi-view renderings are inconsistent in color and lighting, so they can't be directly applied to a mesh. The pipeline:

Back-projection with confidence estimation: renderings are back-projected onto a texture-atlased mesh, producing normalized colors plus confidences derived from view-direction alignment and depth discontinuities, downweighting texels near silhouettes;
Optical-flow registration: flow-fields on the mesh align colors across views, weighted by confidence;
Gradient-domain stitching: advected colors and confidences are fused into a single seamless texture image.

The framework lets image generative models plug directly into geometry-processing pipelines, offering a practical route to automatic texturing of 3D assets.

Original post →

JHU and Adobe open-source system fusing inconsistent multi-view renderings into one seamless texture

More from Multimodal