Research map / Image generation / Text-to-image diffusion
Score Distillation Sampling with Learned Manifold Corrective
Thiemo Alldieck, Nikos Kolotouros, Cristian Sminchisescu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace L_proj with learned manifold corrective L_lmc in SDS | SDS · cats-to-others network training | CLIP-R-Precision: 0.355 -> 0.592 | Improved |
| use LMC-SDS instead of SDS with high guidance omega=100 | SDS · cats-to-others network training | LPIPS: 0.201 -> 0.178 LPIPS, 0.259 -> 0.261 CLIP | Mixed |
| use LMC-SDS instead of DDS on LPIPS | DDS · cats-to-others network training | LPIPS: 0.143 -> 0.178 | Hurt |
| optimize with L_cond only (remove L_lmc) (ablation) | LMC-SDS · optimization-based image synthesis | visual quality (qualitative) | Hurt |
| use high guidance omega in original SDS (ablation) | SDS · image synthesis and DreamFusion | visual quality (qualitative) | Mixed |
| fix noise epsilon only in L_cond while sampling it in L_lmc (ablation) | LMC-SDS · optimization-based editing (10 seeds) | inter-run LPIPS: 0.041 -> 0.158 | Improved |
| fix epsilon in original SDS (ablation) | SDS · optimization-based image synthesis | stability | Hurt |
| use LMC-SDS in DreamFusion with guidance 10 to 20 instead of SDS with omega=100 | DreamFusion · text-to-3D generation | geometry detail (qualitative) | Improved |
| use multi-step SDS (MS-SDS, five denoising steps) instead of SDS | SDS · optimization-based editing | editing effectiveness | No clear effect |
Closest papers by meaning
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.92)
- Identity-preserving Distillation Sampling by Fixed-Point Iterator (CVPR 2025 · similarity 0.90)
- Noise-free Score Distillation (ICLR 2024 · similarity 0.90)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.89)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.89)
- Posterior Distillation Sampling (CVPR 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.
