Mingwei Li

Standard latent VAEs quietly introduce up to 8.5° of surface normal error at object boundaries—TransNormal-2 eliminates this decoding bottleneck on both sides of the pipeline to outperform state-of-the-art geometry models with a 98.6% reduction in supervised normal data.

LoRA's seemingly simple parameter updates can be precisely dissected into layer-specific effects, revealing the underlying mechanisms of its influence on model outputs.