Research map / Image generation / Autoregressive and latent generation
LayerSync: Self-aligning Intermediate Layers
Yasaman Haghighi, Bastien Van Delft, Mariam Hassan, Alexandre Alahi
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 |
|---|---|---|---|
| add LayerSync self-alignment of shallow to deep layers | SiT-XL/2 · ImageNet 256x256 (800 epochs, no CFG) | FID: 8.99 -> 6.87 | Improved |
| add LayerSync instead of Dispersive loss | SiT-XL/2 · ImageNet 256x256 (80 epochs) | FID: 15.95 -> 11.24 | Improved |
| add LayerSync | SiT-B/2 · ImageNet 256x256 (80 epochs) | FID: 36.19 -> 30.00 | Improved |
| add LayerSync | SiT-XL/2 with SDE sampler · ImageNet 256x256 | FID: 8.3 (1400 ep) -> 8.29 (160 ep) | Improved |
| add LayerSync | MTG-Jamendo audio SiT-XL · MTG-Jamendo | FAD (CLAP): 0.251 -> 0.199 | Improved |
| add LayerSync | MDM · HumanML3D | FID: 0.5206 -> 0.4801 | Improved |
| add LayerSync to REPA | SiT-XL/2 + REPA · ImageNet 256x256 (200k iterations) | FID: 7.88 -> 7.01 | Improved |
| use random layer pairings instead of the heuristic layer selection (ablation) | LayerSync SiT-XL · ImageNet 256x256 | FID: 12.24 (std 0.8) vs baseline 17.98 | No clear effect |
| vary lambda from 0.1 to 0.7 (ablation) | LayerSync SiT-B/2 · ImageNet 256x256 | FID: 31.02 to 31.63 | No clear effect |
| add LayerSync, compare 160-epoch LayerSync model vs 1400-epoch baseline (ablation) | SiT-XL/2 · Tiny ImageNet linear probing | avg accuracy: +32.4% avg accuracy | Improved |
| prune layers of a LayerSync-trained model (ablation) | SiT-XL/2 · layer removal | generation quality | Hurt |
| use LayerSync* with CFG scheduling vs REPA* | SiT-XL/2 · ImageNet 256x256 (with CFG, 800 epochs) | FID: 1.49 vs REPA* 1.42 | Mixed |
Later papers that compared with this method
1 later accepted paper reports + LayerSync as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Self-transcendence: Is External Feature Guidance Indispensable for Accelerating Diffusion Transformer Training? (ECCV 2026 · 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 Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.92)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.90)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.90)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.90)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 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.
