Research map / Image generation / Autoregressive and latent generation
REGLUE Your Latents with Global and Local Semantics for Entangled Diffusion
Giorgos Petsangourakis, Christos Sgouropoulos, Bill Psomas, Theodoros Giannakopoulos, Giorgos Sfikas, Ioannis Kakogeorgiou
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 linear PCA with nonlinear semantic compressor for patch features (ablation) | patch-level joint modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 21.4 -> 14.3 | Improved |
| add REPA patch-level alignment to linear patch modeling (ablation) | ReDi · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 21.4 -> 18.8 | Improved |
| add local patch alignment to CLS-only modeling (ablation) | CLS-token joint modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 25.7 -> 15.5 | Improved |
| align only global CLS features without local alignment (ablation) | CLS-token joint modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 25.7 -> 33.7 | Hurt |
| add REPA alignment on top of nonlinear patch modeling (ablation) | nonlinear local modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 14.3 -> 14.1 | Improved |
| add global CLS token and align both local and global (ablation) | nonlinear local modeling plus REPA · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 14.1 -> 13.7 | Improved |
| aggregate last four DINOv2 layers (9-12) before compression (ablation) | nonlinear local modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 14.3 -> 13.3 | Improved |
| include shallow layers 3 and 6 in the aggregation (ablation) | nonlinear local modeling · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 14.3 -> 16.9 | Hurt |
| use multi-layer linear PCA instead of nonlinear compressor (ablation) | stacked baseline components · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 17.4 -> 23.1 | Hurt |
| add KL regularization to the compressor (ablation) | semantic compressor · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 14.3 -> 17.2 | Hurt |
| swap DINOv2-B for DINOv3-B (ablation) | REGLUE · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 12.9 -> 12.3 | Improved |
| swap DINOv2-B for CLIP-L (ablation) | REGLUE · ImageNet 256 conditional (SiT-B/2, 400K, no CFG) | FID: 12.9 -> 18.1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ReDi — ReDi: Efficient Learning-Free Diffusion Inference via Trajectory Retrieval (ICML 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.92)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.91)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.91)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.90)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.90)
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.
