Research map / Architectures and efficiency / Pruning and efficient tuning
Self-Supervised Weight Templates for Scalable Vision Model Initialization
Yucheng Xie, Fu Feng, Ruixiao Shi, Jing Wang, Yong Rui, Xin Geng
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 |
|---|---|---|---|
| use Tucker-based constraint instead of no constraint in pre-training (ablation) | SWEET · ImageNet classification, L5H3 | Top-1 accuracy: 54.51 -> 56.70 | Improved |
| use Tucker constraint instead of Kronecker constraint (ablation) | SWEET · ImageNet classification, L5H3 | Top-1 accuracy: 55.88 -> 56.70 | Improved |
| use Tucker constraint instead of linear factorization (ablation) | SWEET · ImageNet classification, L5H3 | Top-1 accuracy: 55.66 -> 56.70 | Improved |
| add width-wise stochastic scaling during pre-training (ablation) | SWEET · ImageNet classification, L3H4 | Top-1 accuracy: 53.92 -> 55.21 | Improved |
| add width-wise stochastic scaling during pre-training (ablation) | SWEET · ImageNet classification, L3H5 | Top-1 accuracy: 58.65 -> 58.80 | Improved |
| add SwiGLU, RMSNorm and RoPE architectural enhancements (ablation) | SWEET · ImageNet classification, L3H5 | Top-1 accuracy: 57.80 -> 58.80 | Improved |
| replace classification-trained templates (WAVE) with self-supervised Tucker templates (SWEET) | weight-template initialization · ImageNet classification, average over 5 sizes | Top-1 accuracy: 67.19 -> 68.79 | Improved |
| use SWEET instead of WAVE for detection initialization | ViT detectors · COCO detection, average | APbox: 29.55 -> 31.59 | Improved |
| use SWEET instead of WAVE for segmentation initialization | ViT segmenters · ADE20K, average | mIoU: 30.32 -> 33.08 | Improved |
| use SWEET instead of DMAE for generation initialization | ViT generators · ImageNet generation, average | FID: 28.86 -> 26.67 | Improved |
| use SWEET instead of baselines on small fine-grained datasets | ViT L6H6 · Oxford Flower | Top-1 accuracy: 83.18 -> 83.12 | No clear effect |
| use SWEET instead of WAVE for ConvNeXt-v2 initialization | ConvNeXt-v2 atto-L4 · ImageNet classification | Top-1 accuracy: 57.20 -> 57.31 | No clear effect |
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.
- WAVE — WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models (CVPR 2025) · better on 9 of 9 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
- WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models (CVPR 2025 · similarity 0.87)
- Initializing Variable-sized Vision Transformers from Learngene with Learnable Transformation (NeurIPS 2024 · similarity 0.85)
- Rethinking Generative Image Pretraining: How Far Are We From Scaling Up Next-Pixel Prediction? (ICML 2026 · similarity 0.85)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.85)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.85)
- Structured Initialization for Vision Transformers (NeurIPS 2025 · similarity 0.84)
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.
