Research map / 3D vision / 3D generation

Velocity-Space 3D Asset Editing

NeurIPS 2026 · 3D vision › 3D generation · approach: Feed-forward 3D asset generation and editing · 0 citations (OpenAlex)

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.

Proposes VS3D, a training-free, mask-free, inversion-free native-3D editor that modifies the velocity field of a frozen TRELLIS 2.0 rectified-flow pipeline via three modules (RASI, PMG, TAR) to preserve identity while amplifying edits.

Problem

Local editing of 3D assets requires a pretrained rectified-flow 3D generator to produce strong edits on the target region while leaving the rest unchanged; existing editors enforce locality externally via masks, post-hoc merging, or multi-view lifting.

Approach

Intervenes inside the ODE velocity field: learns a per-step, asset-specific unconditional embedding via a source-reconstruction probe to absorb identity leakage (RASI), extrapolates between full and partial Monte-Carlo velocity-difference means to amplify edits (PMG), and uses a twin-forward condition-swap agreement map to blend preserved tokens back to source latents in sparse geometry/material s

Key results

• On the Objaverse-derived editing benchmark, VS3D reaches PSNR 22.51 and LPIPS 0.145, versus 20.40 PSNR and 0.199 LPIPS for the strongest baseline VecSet-Edit.
• VS3D attains DINO-I 0.841 and CLIP-T 0.294, the best on both fidelity measures, while Chamfer Distance (0.034) is second to VoxHammer (0.024).
• Full pipeline runs in 57 s per asset on a single RTX 4090, faster than VoxHammer (394 s), Edit360 (495 s) and Nano3D (143 s).

Surprising findings

• Plain FlowEdit shows severe voxel-occupancy drift on non-edited regions, and adding RASI alone reduces edit strength, requiring PMG to recover it.

Limitations

• Edit quality is upper-bounded by the generative capacity of the frozen TRELLIS 2.0 backbone.
• Texture in non-edited regions is close to but not identical to the source because no hard mask isolates the boundary.
• Per-step RASI optimisation adds cost and hyperparameters (K inner steps, learning rate, early stop).

Experimental setup

Datasets: Objaverse

Benchmarks: Objaverse-derived editing benchmark

Metrics: PSNR SSIM LPIPS DINO-I CLIP-T Chamfer Distance

Baselines: TRELLIS 2.0 VoxHammer Nano3D Instant3dit Edit360 VecSet-Edit

Builds on: FlowEdit TRELLIS 2.0 classifier-free guidance Nano3D

Closest papers by meaning

• 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.87)
• DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation (ICLR 2025 · similarity 0.86)
• PrEditor3D: Fast and Precise 3D Shape Editing (CVPR 2025 · similarity 0.86)
• FlowDrag: 3D-aware Drag-based Image Editing with Mesh-guided Deformation Vector Flow Fields (ICML 2025 · similarity 0.86)
• CTRL-D: Controllable Dynamic 3D Scene Editing with Personalized 2D Diffusion (CVPR 2025 · similarity 0.85)
• FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly (AAAI 2024 · similarity 0.85)