Research map / Image generation / Autoregressive and latent generation
Geometric Decoupling: Diagnosing the Structural Instability of Latent
Yuanbang Liang, Zhengwen Chen, Yu-Kun Lai
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 |
|---|---|---|---|
| switch from normal to OOD prompts | SD3.5 latent geometry · Normal vs OOD prompts | Spearman rho(LC, PHFE): 0.413 -> 0.083 | Hurt |
| switch from normal to OOD prompts | FLUX.1 latent geometry · Normal vs OOD prompts | Spearman rho(LC, PHFE): 0.448 -> 0.269 | Hurt |
| switch from normal to OOD prompts | SD3.5 local scaling · Normal vs OOD prompts | Spearman rho(LS, PHFE): 0.836 -> 0.824 | No clear effect |
| switch from normal to OOD prompts | SD3.5 latent energy · median stats | PHFE latent: 158.506 -> 252.799 | Improved |
| switch from normal to OOD prompts | SD3.5 image output · median stats | image Var(Laplacian x): 0.0120 -> 0.0131 | No clear effect |
| switch from normal to OOD prompts | SD3.5 spectral structure (Chair prompt) · Spectral coupling | Spectral Isolation Score: 4.909 -> 6.453 | Improved |
| switch from normal to OOD prompts | SD3.5 interpolation · Interpolation trajectory | cumulative path length: 5.374 -> 6.243 | Hurt |
| switch from normal to OOD prompts | SD3.5 interpolation · Extremal increments | 95th percentile step increment: 0.453 -> 0.496 | Hurt |
| use ADD-distilled SD3.5 Turbo instead of SD3.5 Base (ablation) | SD3.5 · OOD prompts | Spearman rho(LC, PHFE) under OOD: 0.083 -> 0.156 | Improved |
| use ADD-distilled SD3.5 Turbo instead of SD3.5 Base (ablation) | SD3.5 · LS-PHFE coupling | correlation drop D(LS-PHFE): -1.5% -> -0.6% | No clear effect |
| use LC/PHFE ratio as OOD score instead of raw LC or LS (ablation) | SD3.5 OOD detection · 500 Normal + 500 OOD | AUROC: 0.427 -> 0.816 | Improved |
Closest papers by meaning
- What Secrets Do Your Manifolds Hold? Understanding the Local Geometry of Generative Models (ICLR 2025 · similarity 0.86)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.86)
- Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.86)
- GPLD3D: Latent Diffusion of 3D Shape Generative Models by Enforcing Geometric and Physical Priors (CVPR 2024 · similarity 0.86)
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.86)
- Hyperbolic Geometric Latent Diffusion Model for Graph Generation (ICML 2024 · 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.
