Generating large, navigable 3D scenes from text remains challenging: existing methods either degrade under camera motion or lack object-level editability. DreamAnywhere addressed this with an object-centric pipeline built around a 360° panorama scaffold, asserted to be modular. We present DreamAnywhereRevisited, which keeps the architecture and panorama front end unchanged and replaces every module downstream. A new semantic-inventory stage names, grounds, cross-view-consolidates and filters objects through an explicit keep/reject decision. Accepted objects are reconstructed as individual instances and placed via a physics-constrained optimiser that couples a cross-entropy method to rigid-body simulation, making stability and non-interpenetration first-class objectives. Background depth is re-estimated with Pi3X conditioned on the measured depth so the two align by construction, and the room point cloud is built by direct panorama unprojection with solid-angle-weighted subsampling. Novel viewpoints are synthesised by GEN3C conditioned on point-cloud renderings, replacing score distillation with real multi-view photometric evidence. The resulting scene is composed into an editable 3DGS representation with screen-space ambient occlusion.

We evaluate DreamAnywhereRevisited against three state-of-the-art methods (WorldExplorer, LayerPano3D, DreamScene360) on walk-through video quality and against the conference version DreamAnywhere on its original scenes. We additionally report per-stage measurements over multiple scenes, an ablation of the decision model, a video-diffusion backbone comparison, and a round-trip self-consistency test against the input panorama. DreamAnywhereRevisited demonstrates that individual pipeline modules can be swapped and augmented as the state-of-the-art progresses, improving on the conference version in the cycle-consistency test and on most walk-through video metrics, while remaining modular and object-centric.