Research map / 3D vision / 3D generation
Geometry-Free View Synthesis: Transformers and No 3D Priors
Robin Rombach, Patrick Esser, Björn Ommer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit warping to the implicit model (hybrid) (ablation) | impl.-depth · ACID | NLL: 5.353 -> 5.341 | No clear effect |
| add monocular depth input to the implicit model (ablation) | impl.-nodepth · RealEstate10K | PSNR: 12.16 -> 12.66 | Improved |
| add depth as extra tokens (catdepth) instead of merged codes (ablation) | impl.-depth · RealEstate10K | NLL: 4.836 -> 4.860 | No clear effect |
| warp learnable embeddings instead of fixed VQGAN features (ablation) | expl.-feat · ACID | NLL: 5.657 -> 5.416 | Improved |
| replace warped features with implicit conditioning on depth and source (ablation) | expl.-emb · ACID | NLL: 5.416 -> 5.353 | Improved |
| use a probabilistic transformer instead of a deterministic explicit model (ablation) | expl.-det · RealEstate10K | FID: 66.66 -> 49.15 | Improved |
| take 32 samples and keep the best instead of one sample | impl.-depth · RealEstate10K | PSNR: 12.66 -> 14.47 | Improved |
| take 32 samples and keep the best instead of one sample | impl.-depth · ACID | PSNR: 15.53 -> 16.80 | Improved |
| use a larger VQGAN latent code (f8 instead of f16) (ablation) | VQGAN · reconstruction | PSNR: 21.14 -> 25.23 | Improved |
| evaluate all test pairs instead of the better of two views | SynSin · RealEstate10K | PSNR: 23.03 -> 19.75 | Hurt |
| add 50% larger viewpoint changes to the evaluation | SynSin · RealEstate10K | PSNR: 23.03 -> 16.05 | Hurt |
| add the predicted depth only via the implicit variant without any depth, versus explicit warped images (ablation) | impl.-nodepth · RealEstate10K | NLL: 4.924 -> 4.956 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SynSin — SynSin: End-to-End View Synthesis From a Single Image (CVPR 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Geometry-Aware Deep Network for Single-Image Novel View Synthesis (CVPR 2018 · similarity 0.91)
- GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers (ICLR 2024 · similarity 0.89)
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 2026 · similarity 0.88)
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.88)
- LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.88)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
