UltraTex: Unleashing 2K Multi-View Diffusion for 3D Texturing
UltraTex pushes image-guided 3D texturing to 2048 resolution by cutting redundant multi-view tokens before diffusion work gets expensive.
The paper says current multi-view diffusion texturing usually runs at 512 or 768 resolution, losing fine detail from high-resolution references. UltraTex removes background tokens, uses block-sparse attention on foreground tokens, and adds foreground-aware VAE decoding to avoid artifacts. The authors also built G-buffer TexVerse, a 268,000-plus asset dataset for 2K multi-view training. They report 20.6x-91.1x training speedups and 22.3x-74.6x inference speedups over their baseline on common dataset samples. HF Daily Papers' note
The paper says current multi-view diffusion texturing usually runs at 512 or 768 resolution, losing fine detail from high-resolution references. UltraTex removes background tokens, uses block-sparse attention on foreground tokens, and adds foreground-aware VAE decoding to avoid artifacts. The authors also built G-buffer TexVerse, a 268,000-plus asset dataset for 2K multi-view training. They report 20.6x-91.1x training speedups and 22.3x-74.6x inference speedups over their baseline on common dataset samples. HF Daily Papers' note
score 5