Research map / Image generation / Fast diffusion sampling
Rethinking Score Distillation as a Bridge Between Image Distributions
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.
Reframes score distillation sampling as a dual Schrödinger bridge transport problem and proposes replacing the unconditional source score with text-described corruption prompts, matching VSD quality at SDS cost.
Problem
Score distillation sampling (SDS) and its variants produce artifacts such as oversaturation and oversmoothing; the source distribution of the current optimized image is poorly approximated, and single-step gradients are linear approximations of the transport path.
Approach
Views SDS-family gradients as first-order approximations of a dual diffusion bridge between the current parametric image distribution and the target natural image distribution, then mitigates source distribution mismatch by using a text prompt describing corruptions (e.g., oversaturated, blurry) as the source condition, avoiding VSD's per-iteration LoRA training.
Key results
- On COCO text-to-image, the proposed method reaches zero-shot FID 67.89 vs. SDS 86.02 and VSD 59.22, at 4.48 min per sample vs. VSD 16.02 min.
- In a 37-user study on 31 NeRF prompts, the proposed method's results were preferred in 75.7% of responses against SDS/VSD pairings.
- Full-path bridge sampling with LoRA source estimate lowers FID to 55.65, indicating reducing linear approximation error also helps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use textual source-distribution guidance after an SDS stage instead of SDS alone | Score distillation for image generation · COCO captions (5K) | Zero-Shot FID: 86.02 -> 67.89 | Helped |
| Use textual source-distribution guidance after an SDS stage instead of SDS alone | Score distillation for image generation · COCO captions (5K) | Zero-Shot CLIP FID: 28.39 -> 20.31 | Helped |
| Use textual source-distribution guidance instead of VSD's LoRA estimate | Score distillation for image generation · COCO captions (5K) | Time per Sample (mins): 16.02 -> 4.48 | Helped |
| Use textual source-distribution guidance instead of VSD | Score distillation for image generation · COCO captions (5K) | Zero-Shot FID: 59.22 -> 67.89 | Hurt |
| Replace single-step bridge estimation with full-path PF-ODE simulation | Unconditional source estimate (SDS) · COCO captions | Zero-Shot COCO FID: 86.02 -> 63.31 | Helped |
| Replace single-step bridge estimation with full-path PF-ODE simulation | Textual Bridge source estimate · COCO captions | Zero-Shot COCO FID: 67.89 -> 60.07 | Helped |
| Replace single-step bridge estimation with full-path PF-ODE simulation | LoRA source estimate (VSD) · COCO captions | Zero-Shot COCO FID: 59.22 -> 55.65 | Helped |
| Vary the negative-prompt descriptors (ablation) | Ours · Text-to-image generation | Visual quality | No clear effect |
| Remove the initial SDS stage and use source descriptors from the start (ablation) | Ours · Text-to-image generation | Geometry quality | Hurt |
Surprising findings
- Simply describing corruptions in the negative prompt (e.g., oversaturated, blurry) matches VSD-level quality without training a LoRA.
- Full-path bridge solving introduces additional artifacts and instability in text-to-3D despite improving 2D FID.
Limitations
- Full-path bridge solving does not transfer to text-to-3D and introduces artifacts and instability.
- Results still do not reach the quality and diversity of reverse-process diffusion sampling.
- Painting-to-real requires manual prompt specification; automatic VLM prompts were inconsistent.
Experimental setup
Datasets: MS-COCO validation captions (5K sampled)
Benchmarks: MS-COCO zero-shot FID CLIP FID CLIP similarity on 31 text-to-3D prompts
Metrics: zero-shot FID CLIP FID CLIP similarity time per sample user preference
Baselines: SDS NFSD CSD VSD DDIM SDEdit Fantasia3D Magic3D
Builds on: Score Distillation Sampling Variational Score Distillation Dual Diffusion Implicit Bridges Noise-Free Score Distillation Classifier Score Distillation
Closest papers by meaning
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.90)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.90)
- Taming Mode Collapse in Score Distillation for Text-to-3D Generation (CVPR 2024 · similarity 0.90)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.90)
- DaCapo: Score Distillation as Stacked Bridge for Fast and High-quality 3D Editing (CVPR 2025 · similarity 0.89)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (ICLR 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.