This study proposes an advanced generative framework for creating digital building representations, building upon our previously developed autoregressive Building Digital Cousin (BDC) generator. Previous approaches to building reconstruction have relied heavily on extensive visual data references, which often lead to significant data deficiencies. Our earlier method partially addressed these limitations by incorporating generative modeling based on textual and image targets. However, this approach still suffered from instability during multi-round generation and from limitations associated with manually configured conditioning. This proposal advances the configuration by introducing a simple but novel perceptual evaluation index, Building Mesh Quality Index (BMQI), for building instances. Additionally, an expanded vocabulary for vertex discretization and a latent image-based conditioning mechanism are introduced to refine the previous pipeline. By setting the generation target to LoD2.0 without minor-level details, experiments on the PLATEAU dataset demonstrate the capability of our model to generate a wide range of building models that conform to the image description while maintaining outstanding perceptual quality in more than 99% of cases. An average improvement of 13% in geometric proximity and a BMQI close to 0.99 further demonstrate the effectiveness of our method in addressing previous deficiencies related to data dependency and appearance conformity in urban building reconstruction.