Research map / Image generation / Fast diffusion sampling
Taming Mode Collapse in Score Distillation for Text-to-3D Generation
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 Entropic Score Distillation (ESD), which re-adds an entropy term to score distillation objectives and implements it via classifier-free guidance on VSD, to mitigate Janus artifacts in text-to-3D generation.
Problem
Score distillation for text-to-3D generation suffers from view inconsistency (the Janus artifact), which the paper attributes to mode collapse from maximum-likelihood-seeking objectives.
Approach
Shows SDS and VSD reduce to per-view maximum likelihood because the entropy of the rendered-view distribution is treated as constant; adds an entropy regularization on the rendered image distribution (ESD) and shows its gradient decomposes into camera-conditioned and unconditional scores, so it can be implemented via classifier-free guidance on VSD.
Key results
- ESD achieves a successful generation rate of 55.83% versus 19.17% for VSD and 15.00% for SDS in human evaluation.
- ESD obtains the best FID (235.915) and CLIP distance (0.714) among SDS (291.860, 0.737) and VSD (265.141, 0.725).
- Ablation shows lambda around 0.5 gives the best balance; lambda=0 recovers VSD and reintroduces Janus artifacts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | CLIP distance: 0.737 -> 0.714 | Helped |
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | FID: 291.860 -> 235.915 | Helped |
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | IQ: 4.295 -> 3.135 | Helped |
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | IV: 4.8552 -> 4.0314 | Hurt |
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | IG: 0.123 -> 0.327 | Helped |
| Use ESD instead of SDS | Text-to-3D generation · 120 text prompts | SR: 15.00% -> 55.83% | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | CLIP distance: 0.725 -> 0.714 | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | FID: 265.141 -> 235.915 | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | IQ: 3.149 -> 3.135 | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | IV: 3.5712 -> 4.0314 | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | IG: 0.137 -> 0.327 | Helped |
| Use ESD instead of VSD | Text-to-3D generation · 120 text prompts | SR: 19.17% -> 55.83% | Helped |
Surprising findings
- Directly fitting the camera-unconditioned score via LoRA is unstable (gradient explosion), whereas the CFG-based implementation is stable.
Limitations
- Entropy term is approximated with a CFG-style mixture and a conditional/unconditional score split, adding hyperparameter lambda that trades off quality vs diversity
- Evaluation relies on a proxy inception-style classifier and human judgment over 120 prompts
Experimental setup
Datasets: 120 text prompts from prior work [58]
Benchmarks: 120 text prompts from DreamFusion-style prompt set [58]
Metrics: CLIP distance FID Inception Quality Inception Variety Inception Gain Successful generation Rate
Baselines: SDS (DreamFusion) VSD (ProlificDreamer) Debiased-SDS Perp-Neg
Builds on: Score Distillation Sampling Variational Score Distillation classifier-free guidance DreamFusion
Closest papers by meaning
- Retrieval-Augmented Score Distillation for Text-to-3D Generation (ICML 2024 · similarity 0.90)
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.90)
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.89)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.89)
- Text-to-3D with Classifier Score Distillation (ICLR 2024 · similarity 0.89)
- Consistent Flow Distillation for Text-to-3D Generation (ICLR 2025 · 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.