Research map / Image generation / Fast diffusion sampling
Noise-free 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.
placeholder
Problem
SDS-based text-to-content generation (e.g., text-to-NeRF) produces over-smoothed, over-saturated results and requires large classifier-free guidance scales (e.g., 100), which is not well understood.
Approach
Decomposes the predicted score into condition (CFG difference), domain-correction, and denoising components; identifies the denoising residual minus noise as the source of SDS noise; approximates the domain term (via the unconditional prediction for small t, or a negative-prompt difference for large t) and drops the residual, yielding NFSD that works at CFG 7.5.
Key results
- NFSD with CFG 7.5 produces sharper 2D latent optimization results (e.g., fur, flames, cake frosting) than SDS at CFG 7.5 and comparable or better detail than SDS at CFG 100.
- In text-to-NeRF comparisons using threestudio, NFSD yields more detailed wings, manes and tails than SDS at CFG 100, with ordinary CFG values.
- NFSD matches or exceeds reported results of DreamFusion, Magic3D, Latent-NeRF, Fantasia3D and ProlificDreamer on qualitative comparisons, with no extra diffusion model fine-tuning unlike ProlificDreamer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SDS with NFSD by removing the noisy denoising residual and approximating domain correction | Stable Diffusion 2.1-base latent optimization · 2D image generation | visual quality | Helped |
| Replace SDS with NFSD by removing the noisy denoising residual and approximating domain correction | NeRF optimization · Text-to-NeRF synthesis | visual quality | Helped |
| Increase SDS CFG scale from 7.5 to 100 | SDS · 2D image generation | visual quality | Mixed |
| Use SDS with nominal CFG scale 7.5 instead of NFSD | 2D image generation · 2D image generation | visual quality | Hurt |
| Use SDS instead of NFSD | NeRF optimization · Text-to-NeRF synthesis | visual quality | Hurt |
| Use NFSD instead of ProlificDreamer | Text-to-NeRF synthesis · Text-to-NeRF synthesis | visual quality | No clear effect |
Surprising findings
- Large CFG scales (e.g., 100) commonly used for SDS are not necessary once the denoising residual is removed; NFSD works at CFG 7.5.
- The VSD loss can be shown to approximate the same terms as NFSD, with the extra fine-tuned diffusion model acting as a denoising estimate.
Limitations
- Evaluation is primarily qualitative; no quantitative metrics or benchmarks for NeRF quality are provided
- Decomposition relies on approximations (time threshold t<200, negative-prompt assumption that delta_C for negative prompt equals minus delta_D)
- Low diversity and the Janus (multi-face) problem persist
Experimental setup
Baselines: SDS (CFG 100) SDS (CFG 7.5) DreamFusion Magic3D Latent-NeRF Fantasia3D ProlificDreamer
Builds on: DreamFusion Score Distillation Sampling classifier-free guidance ProlificDreamer DDS
Closest papers by meaning
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.92)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.89)
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.88)
- Text-to-3D with Classifier Score Distillation (ICLR 2024 · similarity 0.88)
- Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (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.