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.
Figures & tables
Figure 1: Overview of Gate3DGen. (A) Data Engine. We render 3D assets from multi-view, select informative anchor views, and apply composite degradations to construct aligned observation-reference pairs (xo,xc) . (B) Gate3DGen. Router Rϕ bypasses inputs classified as clean ( γ=0 ). Otherwise ( γ=1 ), its features h guide the one-step DiT purifier Cθ . xc conditions the frozen 3D generator G . Clean references and spatial losses are used only in training.
Table 1: Degradation-aware routing on synthetic and real images (%). Recall is for Purify . Accuracy is overall ( Purify and Keep ). H=2Recall×Accuracy/(Recall+Accuracy) .
Table 2: Synthetic 2D purification. Fidelity metrics use clean references. Input is unedited, Grad. is Sobel gradient MAE, and IoU is foreground mask overlap. ∗ denotes the commercial closed-source model.
Table 3: Downstream 3D metrics on TRELLIS.2. Input is degraded, and Oracle uses clean references. MD, CD, and F1all are ×10−3 , and F1outer is ×10−1 . PSNR-N evaluates normals. LPIPS compares four-view shaded and albedo renderings. ∗ denotes a commercial closed-source model.
Figure 2: Comparison of images (2D) and Pixal3D outputs (3D). Original (3D) uses unedited inputs. Ours (2D) uses purified images. Boxes mark structure and appearance changes.
Table 4: In-the-wild comparisons against 2D editing baselines. FD-DINOv2 is calculated on multi-view renders per method. VLM scores span 0 – 10 . Purification measures degradation removal in both VLM evaluation and the user study. Pref% reports preferences for Gate3DGen over each baseline. ∗ marks commercial closed-source models.
Table 5: Purifier ablations on our synthetic benchmark. We report 2D image fidelity and downstream 3D metrics from TRELLIS.2. Checkmarks indicate the use of router features ( h ) and spatial losses (region-weighted reconstruction and edge). MD, CD, and F1all are ×10−3 , and F1outer is ×10−1 . Shaded and albedo LPIPS use the corresponding four-view renderings.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Table 6: Router ablations on synthetic evaluation inputs. Variants use an MLP on VAE latents or omit vision LoRA. Recall and accuracy are percentages. Bold marks column maxima.
Figure 3: Oracle-normalized comparison of 2D purification and downstream 3D generation. The radar plot summarizes percentage progress from the observed degraded input (0) toward the matched clean-reference oracle (1) using the metrics reported in Tables 2 and 3 . Negative values indicate deterioration relative to the observed input. Blue labels denote 2D image metrics, green labels denote 3D geometry and normal metrics, and orange labels denote rendered-appearance metrics. Lower-is-better metrics are direction-adjusted so that farther outward is always better.
Table 7: Downstream 3D metrics on Pixal3D. Input and Oracle use degraded and clean-reference images, respectively. MD and CD are ×10−3 . F1all is all-surface F1@10−8 ( ×10−3 ). F1outer is outer-surface F1@10−6 ( ×10−1 ). PSNR-N evaluates rendered normals. LPIPS and albedo LPIPS use four-view shaded and albedo renderings, respectively. Bold and underline mark the best and second-best editing results, excluding Input and Oracle. ∗ denotes a commercial closed-source model.