Research map / Architectures and efficiency / Pruning and efficient tuning
Keep It Frozen: Domain-Routed Conditional Residual Modulation for Multi-Domain Vision Transformers
Ufaq Khan, Umair Nawaz, Massimo Caputo, Muhammad Bilal, Junaid Qadir, Muhammad Haris Khan
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 Residual Modulation Blocks to the frozen ViT (ablation) | DCRM-ViT · FPUS23 | Accuracy (%): 51.4 -> 63.4 | Improved |
| add Residual Modulation Blocks (ablation) | DCRM-ViT · Fetal Planes | Accuracy (%): 74.5 -> 89.3 | Improved |
| add Domain Router for soft domain weights (ablation) | DCRM-ViT · Fetal Planes | Accuracy (%): 78.2 -> 89.3 | Improved |
| add bi-level meta-learning (ablation) | DCRM-ViT · Fetal Planes | Accuracy (%): 83.1 -> 89.3 | Improved |
| add Parameter Synthesizer Network (ablation) | DCRM-ViT · Food101 | Accuracy (%): 92.2 -> 95.7 | Improved |
| add domain-aware attention bias (DAB) (ablation) | DCRM-ViT · FPUS23 | Accuracy (%): 60.1 -> 63.4 | Improved |
| add gating channel unit (ablation) | DCRM-ViT · FPUS23 | Accuracy (%): 60.7 -> 63.4 | Improved |
| add drop-path regularization (ablation) | DCRM-ViT · Food101 | Accuracy (%): 95.2 -> 95.7 | No clear effect |
| use GELU activation component (ablation) | DCRM-ViT · SUN397 | Accuracy (%): 81.1 -> 81.3 | No clear effect |
| increase number of RMB modules up to 12 layers (ablation) | DCRM-ViT · multi-domain classification | performance | Mixed |
| replace LoRA with DCRM-ViT | frozen ViT adaptation · Fpus23 (fine-tuned) | Accuracy (%): 63.0 -> 63.4 | Improved |
| use DCRM-ViT instead of CLIP for zero-shot transfer | zero-shot classification · CIFAR-10 (zero-shot) | Accuracy (%): 71.9 -> 75.2 | Improved |
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.
- AdaptFormer — AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition (NeurIPS 2022) · 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
- Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging (AAAI 2025 · similarity 0.86)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.85)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- DDViT: Double-Level Fusion Domain Adapter Vision Transformer (Student Abstract) (AAAI 2024 · similarity 0.85)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.85)
- Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning (ICML 2025 · similarity 0.85)
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.
