Research map / Image generation / Fast diffusion sampling
Score Distillation via Reparametrized DDIM
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.
Shows that SDS guidance is a high-variance reparametrized DDIM step and proposes Score Distillation via Inversion (SDI), which replaces SDS's random noise with prompt-conditioned DDIM inversion to reduce over-smoothing in text-to-3D generation.
Problem
Score Distillation Sampling (SDS) produces over-smoothed, over-saturated, cartoon-like 3D shapes compared with the quality of the underlying 2D diffusion model, and the reason for this gap was not formally explained.
Approach
Reparametrize DDIM in terms of the single-step denoised estimate x0(t) to show SDS is DDIM with i.i.d. random noise in place of the trajectory-consistent noise; replace that noise with prompt-conditioned DDIM inversion (negative CFG on inversion, positive on forward) to get a better noise estimate, with no extra networks.
Key results
- SDI reaches CLIP Score 33.47 with 10k steps versus 29.81 for SDS and 33.31 for VSD (25k steps), while using 119 min and 39.2GB VRAM.
- On the ISM codebase, SDI improves ImageReward from -0.52 (ISM) to -0.30 at 5k steps with equal VRAM (15.4GB).
- SDI lowers CFG to 7.5 from the typical 100 used by SDS, avoiding over-saturation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace resampled Gaussian noise with DDIM-inversion-inferred noise (ablation) | SDS · Text-to-3D | visual generation quality | Helped |
| Use DDIM inversion to infer the noise term (ablation) | SDI · Text-to-3D | visual generation quality | Helped |
| Use fixed-point iteration to infer the noise term instead of DDIM inversion (ablation) | SDI · Text-to-3D | visual generation quality | No clear effect |
| Use gradient-descent optimization to infer the noise term instead of DDIM inversion (ablation) | SDI · Text-to-3D | visual generation quality | No clear effect |
| Use fixed-point iteration to infer the noise term (ablation) | SDI · Text-to-3D | generation convergence | Hurt |
| Use unconditional DDIM inversion rather than prompt-conditioned inversion (ablation) | SDI · Text-to-3D | visual generation quality | Hurt |
| Set inversion CFG equal to forward CFG (7.5) rather than negative forward CFG (-7.5) (ablation) | SDI · Text-to-3D | inversion numerical error | Hurt |
| Use 10 DDIM inversion steps (ablation) | SDI · Text-to-3D | generation quality and speed | Mixed |
Surprising findings
- DDIM inversion works best when the reverse pass uses negative CFG (-7.5) while the forward pass uses +7.5, which is counter to the reported high inversion error with positive CFG.
- Unconditional inversion (matching ISM) causes over-saturation, while prompt-conditional inversion removes it.
Limitations
- Occasionally produces flat or concave billboard geometry and content drift across views.
- Output quality is capped by the underlying 2D diffusion model, which can produce anomalies such as extra limbs.
- Limited diversity across seeds at coarse level and minimal diversity in human prompts.
Experimental setup
Benchmarks: 43 text prompts for 3D generation (CLIP score, CLIP IQA, ImageReward)
Metrics: CLIP Score CLIP IQA ImageReward divergence rate run time VRAM
Baselines: DreamFusion (SDS) SJC VSD (ProlificDreamer) ESD HiFA Interval Score Matching NFSD
Builds on: DreamFusion (SDS) DDIM Interval Score Matching classifier-free guidance
Closest papers by meaning
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.90)
- Noise Conditional Variational Score Distillation (ICML 2025 · similarity 0.90)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
- Directly Denoising Diffusion Models (ICML 2024 · similarity 0.89)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.89)
- Diffusion Time-step Curriculum for One Image to 3D Generation (CVPR 2024 · similarity 0.89)
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.