SyncDreamer

One photograph in, sixteen aligned viewpoints out: SyncDreamer, an ICLR 2024 spotlight from Yuan Liu and colleagues, reworks single-image novel view synthesis by denoising all target views in one joint diffusion process rather than sampling views independently. A 3D-aware attention mechanism ties features across views at every denoising step, which is what keeps geometry and color stable where a predecessor like Zero123 drifts between generations; the released checkpoint was initialized from zero123-xl and trained on Objaverse renderings. This is research code, not a product: a conda environment, manually fetched checkpoints including a ViT-L CLIP model, an input elevation angle estimated by hand, and a separate NeuS or NeRF fitting stage to turn the sixteen views into an actual mesh. The authors ran it on a 40 GB A100; shrinking the per-step view batch fits under 10 GB of VRAM at a speed penalty. Code and weights are MIT licensed for commercial use. Output quality varies across seeds, and feed-forward successors like InstantMesh finish in seconds, but this remains the reference implementation of synchronized multiview diffusion.