Research map / Image generation / Editing and style transfer
Posterior Distillation Sampling
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 Posterior Distillation Sampling (PDS), an optimization that matches source and target DDPM stochastic latents to edit parametric images (NeRF, SVG) with diffusion priors.
Problem
Score-distillation methods designed for generation lack an explicit identity-preservation term, so editing parametric images (NeRF, SVG) often drifts away from the source content.
Approach
Reformulates DDPM stochastic-latent inversion editing into a score-distillation-style optimization that matches the stochastic latents of the source and target trajectories, so the target's generative posterior stays close to the source's while following the target prompt.
Key results
- NeRF editing: PDS achieves CLIP score 0.2477 vs 0.2280 for IN2N, and 49.33% user preference vs 27.71% for IN2N.
- SVG editing: PDS reaches LPIPS 0.3121 vs 0.4855 (SDS) and 0.5982 (DDS), with CLIP score on par (0.2504).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PDS instead of IN2N | NeRF editing · NeRF editing | CLIP Score: 0.2280 -> 0.2477 | Helped |
| Use PDS instead of IN2N | NeRF editing · NeRF editing | User Preference Rate: 27.71 -> 49.33 | Helped |
| Use PDS instead of SDS | SVG editing · SVG editing | CLIP Score: 0.2606 -> 0.2504 | No clear effect |
| Use PDS instead of SDS | SVG editing · SVG editing | LPIPS: 0.4855 -> 0.3121 | Helped |
| Use PDS instead of SDS | SVG editing · SVG editing | User Preference Rate: 30.83 -> 48.94 | Helped |
Surprising findings
- DDS produces outputs that completely lose the identity of input NeRF scenes, focusing solely on the text prompt.
Limitations
- Editing relies on NeRF reconstruction quality and text prompts that reconstruct the source; requires finetuning Stable Diffusion with DreamBooth for some scenes.
- Evaluation is limited to small sets of scenes and SVGs, with user studies and CLIP score as main metrics.
Experimental setup
Datasets: IN2N scenes LLFF self-captured real scenes VectorFusion SVG inputs
Benchmarks: NeRF editing scenes (13 scenes, 37 pairs) SVG editing (48 pairs)
Metrics: CLIP score LPIPS user preference rate
Baselines: SDS DDS Instruct-NeRF2NeRF Inversion2NeRF
Builds on: Score Distillation Sampling Delta Denoising Score DDPM inversion Instruct-NeRF2NeRF DreamBooth
Closest papers by meaning
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.89)
- DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation (ICLR 2025 · similarity 0.89)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.88)
- PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing (ICLR 2025 · similarity 0.88)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · 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.