Gate3DGen: A Pre-processing Model that Selectively Purifies Real-World Images for 3D Generation
Single-image 3D generators perform well on curated studio photos but fail on real-world images distorted by blur, uneven illumination, or occlusions. Gate3DGen proposes a selective pre-processing pipeline: a degradation-aware router that leaves clean images untouched and a one-step diffusion purifier that repairs corrupted inputs before they reach a frozen 3D generator. The core insight is that improving the observation, rather than retraining the generator, can close most of the fidelity gap while running in a fraction of the time required by generic image editors.
The paper makes a practical contribution by treating robust single-image 3D generation as selective input purification, yielding strong quantitative gains on a synthetic benchmark and a notably fast one-step purifier. However, its real-world claims rest on a small in-the-wild set evaluated with proxy labels and VLM judgments, so the evidence for true out-of-distribution generalization is weaker than the headline numbers suggest.
The routing-plus-purification decomposition is well motivated: a router avoids over-editing clean inputs, while the one-step flow-matching purifier runs at 2.3 s versus 15.6–100.8 s for generic editors. On 1,082 synthetic hold-out cases with ground-truth meshes, the method closes up to 83.1% of the mesh-distance gap to clean-reference inputs and improves both geometry and appearance metrics, whereas all generic editors trade off degradation removal against reconstruction fidelity.
The entire training pipeline depends on synthetic corruptions rendered from 3D assets, creating an unresolved domain gap to real photographs; real-world evaluation is limited to 96 images whose routing labels are assigned by a GPT-5 agent rather than human annotators. On this in-the-wild set, the router does not outperform a trivial always-purify baseline under the proxy metric. More troublingly, the ablation in Table 5 shows that adding the spatial losses increases mesh distance relative to the variant that uses only router features (0.764 vs. 0.755), undercutting the claim that these 3D-oriented constraints are strictly necessary for better geometry.
On the synthetic benchmark the evidence is strong: paired observations and clean references allow precise measurement of 2D fidelity and downstream 3D metrics with two distinct generators. The comparisons against seven image-editing baselines are extensive, though uneven because some are closed-source commercial APIs. The paper is commendably honest that generic editing methods reduce large artifacts while introducing local geometric deviations, and it correctly cautions that image similarity alone is insufficient to predict 3D quality. Still, the analysis lacks statistical tests and is restricted to only two 3D generators, so generalization to other architectures remains speculative.
Architectural details, loss weights, and training hyperparameters are documented in the appendix, and the benchmark builds on publicly named datasets such as TexVerse and DTC. However, the paper does not mention releasing code or model weights, and it relies on proprietary checkpoints such as GPT Image-2.5 Flare and internal assets. Reproduction would require access to the exact synthetic data engine and the 3,961 TexVerse renderings; without these, an independent group cannot recreate the training pairs or verify the claimed one-step inference behavior.
Single-image 3D generation has achieved high fidelity on curated benchmarks, yet degrades sharply on real-world images with uneven illumination, blur, and occlusions. Purifying such inputs requires deciding when to intervene and preserving the visual cues needed for 3D generation, while leaving clean inputs unchanged. We introduce Gate3DGen, a selective pre-processing model that gates image inputs to existing 3D generators without retraining them. A degradation-aware router bypasses inputs classified as clean and extracts features to guide purification of the remaining inputs. Conditioned on these features, a 3D-oriented one-step purifier purifies degraded inputs with spatial losses that encourage preservation of object identity, contours, and textural cues. To provide paired supervision, we construct a synthetic data engine that renders clean views of 3D assets and synthesises composite degradations. On the synthetic benchmark, the purifier improves image fidelity and downstream 3D geometry in only 2.3s on average, reducing TRELLIS.2 mesh distance by 61.5% relative to degraded inputs. VLM evaluation and human preferences on real photographs further support identity preservation and degradation correction. More demos and code are available at https://supra-3d.github.io/gate3dgen.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.