Research map / Image generation / Fast diffusion sampling
DaCapo: Score Distillation as Stacked Bridge for Fast and High-quality 3D Editing
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 DaCapo, a fast score-distillation 3D editing method that stacks diffusion bridges (inversion and generation run simultaneously) with source-target coupling and manifold correction gradients, achieving about 15x speed-up over PDS.
Problem
SDS-based text-driven 3D scene editing is slow (lengthy optimization) and yields low-quality results because editing and preservation objectives conflict across stochastic optimization steps.
Approach
Separates editing and preservation in time with a decreasing diffusion time schedule and reformulates sampling-version DDS as a stacked diffusion bridge approximating DDIB, then adds source-target coupling (alignment loss), classifier-score target estimation, and a manifold correction gradient.
Key results
- DaCapo edits one scene in 35 minutes vs 540 for PDS (about 15x speed-up), within 2500 steps on a single GPU.
- DaCapo CLIP text-image direction similarity 0.2356 vs PDS 0.2477, with CLIP direction consistency 0.9240 (best among methods).
- Evaluated on 8 real scenes yielding 56 source-target pairs; quantitative metrics on 10 edits across two scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DaCapo instead of PDS | 3D scene editing · 10 edits across two scenes | Time (minute): 540 -> 35 | Helped |
| Use DaCapo instead of PDS | 3D scene editing · 10 edits across two scenes | CLIP Direction Consistency: 0.9218 -> 0.9240 | Helped |
| Use DaCapo instead of PDS | 3D scene editing · 10 edits across two scenes | CLIP Text-Image Direction Similarity: 0.2477 -> 0.2356 | Hurt |
| Add source-target coupling (ablation) | DaCapo · 3D scene editing | source content preservation | Helped |
| Add Manifold Correction Gradient (ablation) | DaCapo · 3D scene editing | edit quality | Helped |
Limitations
- Depends on the InstructPix2Pix model and may fail on certain editing cases.
- Quantitative evaluation is limited to CLIP-based metrics on a small subset (10 edits across two scenes); edits are subjective.
Experimental setup
Datasets: InstructNeRF2NeRF scenes PDS scenes
Benchmarks: InstructNeRF2NeRF scenes PDS scenes
Metrics: CLIP text-image direction similarity CLIP direction consistency editing time (minutes)
Baselines: InstructNeRF2NeRF Inversion2NeRF DDS PDS
Builds on: Score Distillation Sampling Delta Denoising Score Dual Diffusion Implicit Bridge Posterior Distillation Sampling Classifier Score Distillation
Closest papers by meaning
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.93)
- DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation (ICLR 2025 · similarity 0.92)
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.89)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.88)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (ICLR 2024 · similarity 0.88)
- Posterior Distillation Sampling (CVPR 2024 · similarity 0.88)
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.