Research map / Image generation / Fast diffusion sampling
Text-to-3D with Classifier Score Distillation
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 the classifier term of classifier-free guidance alone suffices for score-distillation text-to-3D, proposing Classifier Score Distillation (CSD) with annealed negative prompts and editing extensions.
Problem
Score Distillation Sampling for text-to-3D relies on classifier-free guidance, which is usually treated as an auxiliary trick; the role of its classifier component in driving 3D optimization is unclear.
Approach
Use only the classifier score (conditional minus unconditional noise prediction) as the distillation gradient, interpreting it as an implicit noise-aware classifier; this reinterprets negative prompts as dual classifier scores and enables annealed negative weighting and text-driven editing.
Key results
- CSD reaches CLIP Score 78.6 and R@1 81.8% versus 67.5 / 73.1% for DreamFusion and 74.9 / 74.1% for Magic3D.
- In a user study, 59.4% of text-to-3D responses preferred CSD results, and 57.7% in texture synthesis.
- CSD achieves ProlificDreamer-comparable quality in about 1 hour on a single A800 versus 8 hours for ProlificDreamer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SDS with CSD | Text-to-3D generation · Text-to-3D generation | CLIP Score: 67.5 -> 78.6; 74.9 -> 78.6 | Helped |
| Replace SDS with CSD | Text-to-3D generation · Text-to-3D generation | CLIP R-Precision R@1: 73.1 -> 81.8; 74.1 -> 81.8 | Helped |
| Replace SDS with CSD | Text-to-3D generation · Text-to-3D generation | CLIP R-Precision R@5: 90.7 -> 94.8; 91.7 -> 94.8 | Helped |
| Replace SDS with CSD | Text-to-3D generation · Text-to-3D generation | CLIP R-Precision R@10: 97.2 -> 96.3; 96.6 -> 96.3 | Mixed |
| Use CSD instead of the compared methods | Text-to-3D generation · Text-to-3D generation | user preference (%): 30.3 -> 59.4; 10.3 -> 59.4 | Helped |
| Use CSD instead of the compared methods | Texture synthesis · Texture synthesis | user preference (%): 12.4 -> 57.7; 11.0 -> 57.7; 7.9 -> 57.7 | Helped |
| Add a negative prompt to CSD (ablation) | CSD texture synthesis · Texture synthesis | visual quality | Helped |
| Add a negative prompt to CSD (ablation) | CSD texture synthesis · Texture synthesis | text alignment | Hurt |
| Anneal by reducing the negative classifier-score weight (ablation) | CSD texture synthesis · Texture synthesis | visual quality | Helped |
| Anneal by reducing the negative classifier-score weight (ablation) | CSD texture synthesis · Texture synthesis | text alignment | Helped |
| Remove the classifier score from DDS (ablation) | DDS 2D image editing · 2D image editing | editing effectiveness | Hurt |
| Use only the classifier score (ablation) | DDS 2D image editing · 2D image editing | editing effectiveness | Helped |
Surprising findings
- Using the classifier score alone (without the generative term) gives effective 3D optimization, while removing only the classifier term causes optimization to fail.
- Applying CSD to 2D image optimization produces artifacts, unlike the 3D setting.
Limitations
- No distribution-based objective has been formulated to guide the CSD optimization.
- Applying CSD to 2D image optimization yields artifacts.
- Quantitative evaluation is limited to CLIP-based metrics and user studies on a small prompt set.
Experimental setup
Datasets: Objaverse DreamFusion prompt set (81 prompts)
Benchmarks: CLIP Score and CLIP R-Precision on 81 DreamFusion prompts user study on text-to-3D and texture synthesis
Metrics: CLIP Score CLIP R-Precision user preference (%)
Baselines: DreamFusion Magic3D Fantasia3D ProlificDreamer TEXTure
Builds on: DreamFusion Magic3D ProlificDreamer Classifier-free guidance Delta Denoising Score
Closest papers by meaning
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.90)
- Taming Mode Collapse in Score Distillation for Text-to-3D Generation (CVPR 2024 · similarity 0.89)
- Noise-free Score Distillation (ICLR 2024 · similarity 0.88)
- LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching (CVPR 2024 · similarity 0.88)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.88)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (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.