SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion

SynCity 3000, from the University of Oxford’s Visual Geometry Group with support from Meta Research, generates globally coherent, layout-controlled 3D worlds from text. Its two-stage pipeline first uses Flux ControlNet and a MultiDiffusion-inspired overlapping-window process to create a dimetric 2D scene template, then fine-tunes the TRELLIS image-to-3D model for convolutional inference over sliding 3D windows. The model produces a coarse voxel structure, DINOv2-conditioned structured latents, and ultimately 3D Gaussian Splats that can scale beyond object-sized outputs. To overcome the shortage of scene-scale training data, the authors procedurally place Objaverse-XL assets on randomized terrains, producing 320k synthetic samples for fine-tuning TRELLIS’s sparse-structure and latent transformers for 260k and 660k steps on 2× NVIDIA RTX A6000 GPUs. On template faithfulness, the full method reaches 0.5247 SSIM and 14.4137 PSNR, outperforming off-the-shelf TRELLIS and alternatives including TripoSG and Hunyuan3D-2.1. In a 27-person user study, SynCity 3000 achieved a 78.6% overall scene-preference win rate against TRELLIS using its template, while users unanimously preferred its flexible layout control over the grid-based SynCity. Layout instructions in the experiments were generated with ChatGPT 5 Instant. Remaining limitations include dimetric-perspective requirements, reduced texture fidelity from window averaging, occasional duplicated structures, and a cartoonish appearance inherited partly from FLUX and Objaverse-XL.