Research map / Image generation / Fast diffusion sampling
LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations
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 LGDM, which guides sampling of a frozen Stable Diffusion v1.5 with perceptual manifold guidance (PMG) and regresses NR-IQA scores from multi-scale, multi-timestep U-Net "diffusion hyperfeatures".
Problem
No-reference image quality assessment (NR-IQA) with generalizable features; little exploration of perceptual consistency in latent diffusion models for NR-IQA.
Approach
Use an unconditionally pretrained latent diffusion model as a frozen feature extractor, steering each DDIM step with a likelihood term against the input image plus a perceptual-feature guidance term (PMG), then regress quality from aggregated U-Net hyperfeatures across layers and timesteps.
Key results
- LGDM-ψSDv1.5 reaches PLCC 0.940 and SRCC 0.908 on LIVEC, surpassing LoDA (0.899/0.876).
- On FLIVE, LGDM-ψSDv1.5 achieves PLCC 0.812 and SRCC 0.705 versus prior best of 0.733/0.645 (RE-IQA).
- On KADID, LGDM-ψSDv1.5 reaches PLCC 0.961 and SRCC 0.958, outperforming LoDA's 0.936/0.931.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LGDM-ψSDv1.5 instead of the strongest listed baseline | NR-IQA methods · LIVEC | PLCC: 0.913 -> 0.940 | Helped |
| Use LGDM-ψSDv1.5 instead of the strongest listed baseline | NR-IQA methods · FLIVE | SRCC: 0.645 -> 0.705 | Helped |
| Use Stable Diffusion 1.5 as the perceptual guidance model instead of no perceptual guidance (ablation) | LGDM · LIVEC | PLCC: 0.853 -> 0.940 | Helped |
| Use Stable Diffusion 1.5 as the perceptual guidance model instead of no perceptual guidance (ablation) | LGDM · FLIVE | SRCC: 0.672 -> 0.705 | Helped |
| Use BRISQUE features for perceptual guidance instead of no perceptual guidance (ablation) | LGDM · FLIVE | SRCC: 0.672 -> 0.598 | Hurt |
| Remove the image-consistency update by setting ζ1=0 (ablation) | LGDM with Stable Diffusion 1.5 guidance · LIVEC | PLCC: 0.940 -> 0.901 | Hurt |
| Increase sampling timesteps from 1 to 10 (ablation) | LGDM-ψSDv1.5 · FLIVE | SRCC: 0.624 -> 0.705 | Helped |
| Increase sampling timesteps from 10 to 50 (ablation) | LGDM-ψSDv1.5 · FLIVE | SRCC: 0.705 -> 0.711 | Helped |
| Use Stable Diffusion 1.5 instead of Stable Diffusion 2 (ablation) | LGDM · LIVEC | SRCC: 0.882 -> 0.908 | Helped |
| Use Stable Diffusion 1.5 instead of Stable Diffusion 2.1 (ablation) | LGDM · FLIVE | SRCC: 0.681 -> 0.705 | Helped |
Surprising findings
- Guidance with weaker perceptual features such as BRISQUE reduces performance relative to no perceptual guidance on several datasets.
- Newer Stable Diffusion versions (v2.0, v2.1) yield lower IQA correlation than SD v1.5.
Limitations
- Inference is slow relative to feed-forward IQA models (about 9.7 s for 10 steps).
- Performance depends on the choice of perceptual feature model psi_p; weaker features can hurt results.
- Theoretical manifold claims rest on assumptions and proof sketches.
Experimental setup
Datasets: LIVEC KonIQ FLIVE SPAQ LIVE CSIQ TID2013 KADID AGIQA-1K AGIQA-3K
Benchmarks: LIVEC KonIQ FLIVE SPAQ LIVE CSIQ TID2013 KADID AGIQA-1K AGIQA-3K
Metrics: PLCC SRCC
Baselines: ILNIQE BRISQUE WaDIQaM DBCNN TIQA MetaIQA P2P-BM HyperIQA TReS MUSIQ RE-IQA LoDA
Builds on: DDIM Tweedie's formula diffusion hyperfeatures Stable Diffusion v1.5 Re-IQA posterior sampling guidance
Closest papers by meaning
- What Matters When Repurposing Diffusion Models for General Dense Perception Tasks? (ICLR 2025 · similarity 0.88)
- Alias-Free Latent Diffusion Models: Improving Fractional Shift Equivariance of Diffusion Latent Space (CVPR 2025 · similarity 0.87)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.87)
- Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.87)
- FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance (NeurIPS 2026 · similarity 0.87)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.86)
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.