Research map / Image generation / Fast diffusion sampling
Devil is in the Details: Density Guidance for Detail-Aware Generation with Flow Models
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.
Analyzes latent-scaling "Prior Guidance" via a score alignment condition and proposes Density Guidance (ODE/SDE modifications) for exact log-density control of detail in flow/diffusion sampling.
Problem
Image likelihood correlates negatively with perceptual detail, but existing guidance (latent scaling, score scaling) gives no explicit control over sample log-density, and it is unclear when latent scaling monotonically changes detail.
Approach
Introduces score alignment, a tractable condition (checkable via augmented sensitivity ODEs) explaining when scaling the latent code monotonically changes log-density; then derives a modified generative ODE (Density Guidance) that enforces a prescribed log-density evolution by rescaling the score, extended to SDEs via score-orthogonal noise projection.
Key results
- Negative log-density correlates with PNG compressed size at up to 84% on EDM2, versus 29% for FLIPD.
- Augmented sensitivity equations estimate score alignment with 98.8% correlation to ground truth on CIFAR-10.
- Score alignment holds for 97% of latent codes on CIFAR-10 and 100% on EDM2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use negative log-density instead of FLIPD as a measure of image detail | EDM2 · ImageNet 512×512 | correlation with PNG size: 84% vs 29% | Helped |
| Use augmented sensitivity equations to estimate score alignment without evaluating the score directly | VP-SDE diffusion · CIFAR-10 | correlation with ground-truth score alignment: 98.8% | Helped |
| Apply Prior Guidance by scaling latent codes | VP-SDE diffusion · CIFAR-10 | score alignment rate: 97% | Helped |
| Apply Prior Guidance by scaling latent codes | EDM2 · ImageNet512 | score alignment rate: 100% | Helped |
| Use explicit quantile matching to guide density | Density Guidance · CIFAR-10 | correlation between target and obtained log-density: above 99% | Helped |
| Use Density Guidance rather than Prior Guidance | Image generation with EDM2 · EDM2 sampling | sample quality | No clear effect |
| Inject stochasticity at selected stages during density-guided sampling | Stochastic Density Guidance · Generated samples | variation | Helped |
| Scale latent codes in settings where score alignment fails | Prior Guidance · Two-dimensional Gaussian mixture | log-density: 1.22 | Mixed |
Surprising findings
- Score alignment does not hold in general; a 2D Gaussian mixture counterexample shows latent scaling can decrease log-density for some codes and increase it for others.
- Log-density correlates more strongly with image detail at intermediate time (logSNR=1) than at t≈0.
Limitations
- Exact density control requires knowing or estimating the score; quantile-based targets need empirical estimation of log-density distributions.
- Score alignment can fail in general, so Prior Guidance is not guaranteed monotonic.
- Stochastic extension only approximately satisfies the log-density constraint for finite sampling steps.
Experimental setup
Datasets: CIFAR-10 ImageNet 512x512
Benchmarks: CIFAR-10 ImageNet 512x512
Metrics: correlation with PNG size correlation of log-density with desired value correlation of estimated and ground-truth score alignment
Baselines: Prior Guidance FLIPD Auto-guidance Score scaling
Builds on: Prior Guidance probability flow ODE continuous normalizing flows FLIPD
Closest papers by meaning
- Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models (NeurIPS 2025 · similarity 0.89)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.88)
- Theoretical insights for diffusion guidance: A case study for Gaussian mixture models (ICML 2024 · similarity 0.88)
- Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction (ICLR 2025 · similarity 0.87)
- Guiding a Diffusion Model with a Bad Version of Itself (NeurIPS 2024 · similarity 0.87)
- Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 2024 · similarity 0.87)
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.