Research map / Architectures and efficiency / Attention and transformers
RGB No More: Minimally-Decoded JPEG Vision Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train ViT-Ti directly on JPEG DCT with DCT augmentation rather than RGB | ViT-Ti · ImageNet | training pipeline throughput: +39.2 % | Helped |
| Train ViT-Ti directly on JPEG DCT rather than RGB | ViT-Ti · ImageNet | evaluation pipeline throughput: +17.9 % | Helped |
| Train ViT-Ti directly on JPEG DCT rather than RGB | ViT-Ti · ImageNet validation | accuracy: 74.1 -> 75.1 | Helped |
| Use mixed precision for JPEG-S | ViT-S · ImageNet | training pipeline throughput: +36.2 % | Helped |
| Use grouped rather than separate or concatenated embedding (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.7 -> 76.5; 71.3 -> 76.5 | Helped |
| Add sub-block conversion to grouped embedding (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.6 -> 76.5 | Helped |
| Remove sub-block-based augmentation (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.6 -> 75.1 | Helped |
| Remove block-wise and sub-block-based augmentations (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 72.6 vs 74.6 | Hurt |
| Remove special-case and subsequent augmentation groups (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 73.3 vs 74.6 | Hurt |
| Remove frequency-based and subsequent augmentation groups (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.2 vs 74.6 | Hurt |
| Remove DC-based and subsequent augmentation groups (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.6 -> 75.5 | Helped |
| Use the selected augmentation subset instead of all augmentation groups (ablation) | ViT-S with JPEG DCT input · ImageNet validation | accuracy: 74.6 -> 76.5 | Helped |
Closest papers by meaning
- Faster Neural Networks Straight from JPEG (NeurIPS 2018 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.86)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.86)
- Structured Initialization for Vision Transformers (NeurIPS 2025 · similarity 0.86)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.86)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 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.