Research map / Vision-language and multimodal / Cross-modal learning
Synesthesia via Direct Latent Augmentation: Bypassing the Decode-Encode Loop for Cross-Modal Distillation
Cristian Sbrolli, Nicolas Michel, Matteo Matteucci, Toshihiko Yamasaki
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 |
|---|---|---|---|
| train teacher on undecoded latents instead of decoded raw synthetic data | multimodal teacher · Caltech101 | Top-1 accuracy: +7.39% | Improved |
| add synthetic latents (3D + audio) to image (ablation) | Image-only teacher · Caltech101 | Top-1 accuracy: 55.10 -> 66.47 | Improved |
| add 3D latents to image (ablation) | Image-only teacher · ImageNet-100 | Top-1 accuracy: 61.57 -> 70.03 | Improved |
| add audio latents to image (ablation) | Image-only teacher · ImageNet-100 | Top-1 accuracy: 61.57 -> 65.76 | Improved |
| add audio to Image+3D (full three modalities) (ablation) | Image + 3D teacher · ImageNet-100 | Top-1 accuracy: 70.03 -> 71.59 | Improved |
| use Perceiver resampler for raw decoded modalities (ablation) | Raw teacher · multiple benchmarks | Top-1 accuracy: at most 1.13 pp | Improved |
| use MLP projection head instead of L2 matching (CLS target) (ablation) | LUGPI student · ImageNet-100 | Top-1 accuracy: 65.70 -> 67.04 | Improved |
| predict auxiliary-modality summaries instead of teacher CLS (ablation) | CLS-Pred student · ImageNet-100 | Top-1 accuracy: 67.04 -> 68.54 | Improved |
| use DLA latent teacher instead of raw teacher (ablation) | LUGPI student · ImageNet-100 | Top-1 accuracy: 62.31 -> 65.70 | Improved |
| use DLA latent teacher instead of raw teacher (ablation) | MESSy student · ImageNet-100 | Top-1 accuracy: 64.03 -> 68.54 | Improved |
| use MESSy instead of LUGPI distillation | student · Caltech101 | Top-1 accuracy: +3.40% | Improved |
| use MESSy instead of LUGPI distillation | student · CIFAR-100 linear probe | accuracy: 42.4 -> 48.1 | Improved |
Closest papers by meaning
- Efficient Modality Translation via Arbitrary Conditioning and Wasserstein Regularization (AAAI 2026 · similarity 0.87)
- Multimodal Variational Autoencoder: A Barycentric View (AAAI 2025 · similarity 0.86)
- Unity by Diversity: Improved Representation Learning for Multimodal VAEs (NeurIPS 2024 · similarity 0.86)
- A Closer Look at Multimodal Representation Collapse (ICML 2025 · similarity 0.85)
- Towards General Modality Translation with Contrastive and Predictive Latent Diffusion Bridge (NeurIPS 2025 · similarity 0.85)
- Multi-Modal Latent Variables for Cross-Individual Primary Visual Cortex Modeling and Analysis (AAAI 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.
