Learning 3D Editing without Paired Supervision via Generative Prior Distillation

Hao Wen; Weibin Yun; Hongxing Fan; Haotian Lu; Rui Chen; Zehuan Huang; Lu Sheng

cs.CVSep 4, 2026

Learning 3D Editing without Paired Supervision via Generative Prior Distillation

Authors: Hao Wen, Weibin Yun, Hongxing Fan, Haotian Lu, Rui Chen, Zehuan Huang, Lu Sheng

Organizations: Beihang University, China · Beihang University, China and Beijing Key Laboratory of Intelligent Creative Content Generation and Immersive Experience, China · Beihang University, China and VAST, China

Abstract

Instruction-guided 3D editing is essential for interactive content creation, yet it faces a significant bottleneck: the severe scarcity of high-quality paired training data. Existing approaches attempt to bypass this by either relying on slow test-time optimization or training on pseudo-pairs constructed via complex pipelines, which often introduce structural drift and geometric artifacts. In this paper, we propose a novel framework that learns feed-forward 3D editing without paired 3D supervision via Generative Prior Distillation. Instead of relying on ground-truth 3D pairs, our core idea is to distill visual, semantic, and geometric knowledge from powerful foundation models directly into a 3D editing model. Specifically, through a differentiable rendering pipeline, we supervise the 3D representation using two complementary signals: a 2D visual prior from an image editing model at the main editing view, and a semantic prior from a Vision-Language Model at novel views to ensure strict instruction following and source identity preservation. Crucially, to address the geometric collapse and multi-view inconsistencies inherent in 2D projection supervision, we introduce a 3D-aware Distribution Matching regularization. Acting as a geometric prior, this term operates in the 3D latent space, constraining the edited output to remain within the manifold of realistic 3D assets defined by a pretrained image to 3D teacher model. Extensive experiments demonstrate that our method achieves superior instruction fidelity and cross-view consistency, significantly outperforming state-of-the-art baselines. Our project is available at: https://github.com/thiamine128/PriorEdit3D.

Figures & tables

Figure 1. Our method learns feed-forward 3D editing without paired 3D supervision by distilling generative priors from foundation models, achieving high-fidelity instruction-guided editing with cross-view consistency.

Paradigm	Representative Methods	Inv.- free	Native 3D	No Paired 3D Data	Prior Distill.	Mask- free
Optimization	Instruct-N2N, Vox-E	✗	✗	✓	✓	✓
2D Lifting	Tailor3D, EditP23	✓	✗	✓	✗	✓
Training-free	VoxHammer, Nano3D	✗	✓	✓	✗	✗
Supervised	3DEditFormer, ShapeUP	✓	✓	✗	✗	✓
Unpaired Distillation	PriorEdit3D (Ours)	✓	✓	✓	✓	✓

Table 1. Comparison of SOTA 3D editing paradigms. Our framework is the first unpaired approach to achieve inversion-free and mask-free editing in native 3D space. This is accomplished by distilling generative priors rather than relying on pseudo-pairs.

Figure 2. Automated dataset pipeline. Objaverse objects are rendered, edited via Gemini3-Flash instructions and Qwen-Image-Edit, and filtered using Qwen3-VL 32B and SSIM.

Figure 3. Training framework of PriorEdit3D. Given a source 3D latent and an instruction-conditioned edited image, the editor predicts an edited UniLat3D latent, which is rendered and optimized with pixel-level 2D supervision, VLM-based semantic/view feedback, and DMD-based 3D prior regularization.

Figure 4. Qualitative comparison between our method and SOTA methods, including EditP23 ( Bar-On et al., 2025 ) , Instant3DiT ( Barda et al., 2025 ) , 3DEditFormer ( Xia et al., 2026 ) , VoxHammer ( Li et al., 2026 ) , and Nano3D ( Ye et al., 2026 ) .

	Edited Image Alignment			Overall 3D Quality		Condition Alignment
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	FVD $\downarrow$	CLIP-T $\uparrow$	LLM-Id (%) $\uparrow$	LLM-Inst (%) $\uparrow$	Runtime
PriorEdit3D
EditP23 ( Bar-On et al., 2025 )	16.74	0.87	0.19	184.51	286.72	0.29	82.52	64.71	18 s
Instant3DiT ( Barda et al., 2025 )	15.45	0.86	0.20	229.47	328.24	0.23	40.09	25.22	20 s
3DEditFormer ( Xia et al., 2026 )	18.89	0.89	0.15	167.85	220.91	0.28	86.31	53.39	74 s
VoxHammer ( Li et al., 2026 )	16.87	0.87	0.17	179.22	359.50	0.27	76.54	51.74	133 s

Table 2. Main quantitative comparison across different evaluation datasets. We report edited-image alignment, overall 3D quality, condition alignment, and runtime per edit on a single A100 GPU.

Figure 5. Qualitative results on ABO dataset.

Figure 6. Qualitative results on GSO dataset.

Figure 7. Qualitative ablation results. We compare our full method with variants w/o VLM, w/o DMD, w/o Pixel, w/o both VLM and DMD, and direct conditional generation (UniLat3D). Red circles highlight failures: removing VLM degrades non-edited views; removing DMD causes geometric inconsistencies; removing pixel supervision weakens edit fidelity; removing both VLM and DMD causes severe collapse; and UniLat3D fails to preserve source geometry. Our full model ensures accurate edits and 3D consistency.

Loss Terms			Edited Image Align			Overall 3D Quality		Condition Alignment
Pixel	DMD	VLM	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	FVD $\downarrow$	CLIP-T $\uparrow$	LLM-Id (%) $\uparrow$	LLM-Inst (%) $\uparrow$
✓	✗	✗	24.00	0.94	0.12	78.90	307.58	0.30	88.35	80.38
✓	✗	✓	23.21	0.93	0.12	75.13	211.08	0.31	90.10	86.92
✓	✓	✗	24.45	0.94	0.12	73.67	193.44	0.30	90.92	83.46
✗	✓	✓	20.99	0.91	0.14	94.11	197.79	0.30	88.29	86.92
✓	✓	✓	24.37	0.94	0.12	71.96	195.77	0.31	93.13	86.92

Table 3. Ablation study of loss terms. We compare pixel reconstruction loss, VLM loss, DMD loss, and additional leave-one-out ablations.

VLM Backbone	Edited Image Align			Overall 3D Quality		Condition Alignment
VLM Backbone	PSNR (M) $\uparrow$	SSIM (M) $\uparrow$	LPIPS (M) $\downarrow$	FID $\downarrow$	FVD $\downarrow$	CLIP-T $\uparrow$	LLM-Id (0–100) $\uparrow$	LLM-Inst (%) $\uparrow$
Qwen3-VL-2B	24.26	0.94	0.12	74.34	195.90	0.30	91.50	85.77
Qwen3-VL-4B	24.37	0.94	0.12	71.96	195.77	0.31	93.13	86.92

Table 4. Impact of VLM backbone on 3D editing performance.

Prompt Strategy	Edited Image Align			Overall 3D Quality		Condition Alignment
Prompt Strategy	PSNR (M) $\uparrow$	SSIM (M) $\uparrow$	LPIPS (M) $\downarrow$	FID $\downarrow$	FVD $\downarrow$	CLIP-T $\uparrow$	LLM-Id (0–100) $\uparrow$	LLM-Inst (%) $\uparrow$
Edit fidelity only	23.10	0.93	0.12	80.28	211.90	0.31	90.21	83.85
Identity only	24.33	0.94	0.12	71.77	187.01	0.30	91.94	82.31
Full	24.37	0.94	0.12	71.96	195.77	0.31	93.13	86.92

Table 5. Impact of VLM prompt strategy on 3D editing performance.

Figure 10. Representative failure cases.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12. Qualitative results with samples in Edit3D-Bench.

Component	Parameter	Value
Model architecture
Denoiser	Model	SparseStructureFlowModel
Auxiliary model	Model	SparseStructureFlowModel; architecture identical to the denoiser
Both	Sparse resolution; patch size	16; 1
Both	Input/output channels	32 / 32
Both	Model/conditioning channels	1280 / 1280

Appendix

Table 6. Detailed model and training configuration used in our experiments.

Setting	Forward	Backward	Peak Memory
Fake Model (UniLat3D)	6.498 s	1.061 s	–
Pixel	7.782 s	1.246 s	38 GB
Pixel + DMD	73.006 s	11.730 s	43 GB
Pixel + DMD + VLM	73.252 s	11.878 s	59 GB
Inference	7 s	–	13 GB

Appendix

Table 7. Training-time computational cost and peak GPU memory under different supervision settings.

3D Shape Editing

3D shape editing changes the geometry, appearance, or parts of digital objects using spatial controls, prompts, images, or generative models.

See all 53 papers

Feedforward 3D Editing Learns from Semantic-Part Transformation
May 26, 2026Jiawei Weng, Saining Zhang, Zhenxin Diao +43D Asset Generation 3D Asset Editing
Tsinghua University, China
Vinedresser3D: Agentic Text-guided 3D Editing
Feb 23, 2026Yankuan Chi, Xiang Li, Zixuan Huang +1Text-Guided 3D Scene Editing 3D Asset Editing
The Hong Kong University of Science and Technology · University of Illinois Urbana-Champaign
Omni-3DEdit: Generalized Versatile 3D Editing in One-Pass
Mar 18, 2026Chen Liyi, Wang Pengfei, Zhang Guowen +23D Scene Editing 3D Shape Editing
The Hong Kong Polytechnic University

May 26, 2026J/K move · Enter open · S save

Feedforward 3D Editing Learns from Semantic-Part Transformation

Jiawei Weng, Saining Zhang, Zhenxin Diao, Peishuo Li+3

Tsinghua University, China

3D Asset Generation 3D Asset Editing 3D Shape Editing

Open paper AskPDF arXiv

3D editing is a fundamental capability for scalable 3D content creation. While image editing has rapidly evolved toward large-scale feedforward generative paradigms, 3D AI generation remains dominated by training-free editing pipelines. A central challenge of feedforward 3D editing lies in the lack of high-quality paired supervision. Editable 3D assets require simultaneous preservation of geometry, multi-view consistency, structural coherence, and localized edit controllability. Existing 3D editing datasets often rely on independently generated assets, image-mediated reconstruction or narrow edit taxonomies, leading to inaccurate localization, weak preservation, blurred edit boundaries, and limited semantic consistency. In this work, we introduce a new perspective: scalable feedforward 3D editing should be learned from semantic-part transformations. Based on this insight, we propose Pxform, a high-quality 3D editing dataset with over 100K consistent before/after editing pairs across seven edit types. Instead of treating objects as unstructured shapes, our pipeline grounds edits directly in semantic 3D parts. Built upon Pxform, we further propose PartFlow, a feedforward 3D editing network that injects source-aware latent control into pretrained 3D generative priors. PartFlow introduces mask-aware velocity preservation and render-space consistency supervision to jointly improve edit fidelity and source preservation, while requiring no 3D edit mask during inference. Extensive experiments demonstrate that high-quality semantic-part supervision substantially improves scalable 3D editing, enabling PartFlow to achieve state-of-the-art performance on both geometric and appearance editing benchmarks.

Abstract

3D Shape Editing

Explore similar work

Feedforward 3D Editing Learns from Semantic-Part Transformation

Vinedresser3D: Agentic Text-guided 3D Editing

Omni-3DEdit: Generalized Versatile 3D Editing in One-Pass