Texture-Resolution Confounding in Image-to-3D GLB Exports: A Five-Object Paired Audit of Geometry, File Size, and Local Browser Processing
Year: 2026
Type: preprint
Abstract: Description Comparisons of single-image and multi-image 3D reconstruction can be misleading when differences in output settings are mistaken for improvements in reconstruction performance. This research investigates texture-resolution confounding, geometric characteristics, file size, and browser processing efficiency in image-to-3D asset generation. The study evaluates ten GLB models reconstructed from five objects in the YCB benchmark dataset: a cracker box, mustard bottle, mug, power drill, and banana. Each object was reconstructed using one photograph and three photographs, allowing paired comparisons across both conditions. The original single-image models used 2048 × 2048 base-color textures, while the three-image models used 1024 × 1024 textures. After normalizing texture resolution to 1024 × 1024, average GLB file sizes were 2.248 MiB and 2.194 MiB, respectively, representing a difference of approximately 2.4%. This demonstrates that texture resolution can substantially influence apparent differences in exported asset size. A geometric audit examined mesh connectivity and boundary-edge characteristics. Additionally, a simplified local WebGL2 prototype was used to investigate asset-processing performance. Average measured processing times were 32.80 ms for the single-image models and 23.28 ms for the three-image models. However, the small sample size and differences between the original exports prevent these observations from establishing a causal advantage for either reconstruction condition. The findings highlight the importance of controlling texture resolution, export configuration, and evaluation methodology when benchmarking image-to-3D reconstruction systems. The study offers a practical framework for evaluating GLB assets and identifying experimental confounders, while acknowledging that calibrated reconstruction-accuracy measurements remain necessary for stronger conclusions. Keywords: Image-to-3D Reconstruction, 3D Asset Generation, GLB, glTF, Texture Resolution, WebGL2, Geometric Analysis, YCB Dataset, 3D Model Optimization, Reproducibility. Author: Eishal Research Project: Ashes AI Publication Type: Research Preprint Year: 2026
Author: Eishal
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237345311
Cites:
Cited by:
Related to:
