Research map / 3D vision / 3D generation
Cycle3D: High-quality and Consistent Image-to-3D Generation via Generation-Reconstruction Cycle
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Cycle3D cyclically couples a frozen Stable Diffusion 1.5 denoiser with a feed-forward Gaussian-splatting reconstruction model during multi-step diffusion to produce higher-quality, more multi-view-consistent image-to-3D objects.
Problem
Two-stage large reconstruction pipelines for image-to-3D produce low-quality and multi-view-inconsistent images from multi-view diffusion models, leading to geometric artifacts and blurry textures in the reconstructed 3D content.
Approach
Within the denoising loop, predicted clean latents are reconstructed into 3D Gaussians, re-rendered from the same views, and used to form the next sampling state (generation-reconstruction cycle), with reconstruction-model features interacting with 2D diffusion encoder features and reference-view attention injection for consistency.
Key results
- On the test set, Cycle3D reaches PSNR 20.2452, SSIM 0.8729, LPIPS 0.1117 and CLIP similarity 0.8238, beating LGM (18.6909 PSNR, 0.7990 CLIP).
- On GSO, Cycle3D reaches PSNR 21.4841 vs. 19.4269 for LGM and contextual distance 0.8346 vs. 0.9314.
- Ablation: removing feature interaction and reference-view injection drops PSNR from 20.2452 to 19.2491.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature interaction between the 2D diffusion model and reconstruction model (ablation) | Cycle3D · test dataset | PSNR: 19.2491 -> 20.0198 | Helped |
| Add reference-view feature injection during 2D diffusion denoising (ablation) | Cycle3D with feature interaction · test dataset | CLIP: 0.8045 -> 0.8238 | Helped |
| Use feature interaction and reference-view injection (ablation) | Cycle3D · test dataset | Contextual: 1.6399 -> 1.6031 | Helped |
| Use Cycle3D instead of LGM | image-to-3D generation · test dataset | PSNR: 18.6909 -> 20.2452 | Helped |
| Use Cycle3D instead of LGM | image-to-3D generation · GSO | PSNR: 19.4269 -> 21.4841 | Helped |
Limitations
- Limited to object-level generation; cannot extend to scene-level due to lack of large-scale 3D scene datasets.
- Generation takes about 25 seconds per object, slower than purely feed-forward reconstruction.
Experimental setup
Datasets: G-objaverse RealFusion15 Make-It-3D dataset GSO
Benchmarks: GSO (50 objects) RealFusion15 + Make-It-3D collected real-world test set
Metrics: PSNR SSIM LPIPS CLIP similarity contextual distance
Baselines: DreamGaussian Wonder3D One-2-3-45 TriplaneGaussian OpenLRM LGM
Builds on: Large Gaussian Model (LGM) Stable Diffusion 1.5 Zero-1-to-3 style multi-view diffusion 3D Gaussian Splatting DDIM
Closest papers by meaning
- Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (CVPR 2025 · similarity 0.92)
- Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models (ICLR 2025 · similarity 0.92)
- One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion (CVPR 2024 · similarity 0.91)
- Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image (NeurIPS 2024 · similarity 0.91)
- DMV3D: Denoising Multi-view Diffusion Using 3D Large Reconstruction Model (ICLR 2024 · similarity 0.91)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.90)
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.