Research map / Vision-language and multimodal / CLIP and zero-shot
Multimodal Dataset Distillation Made Simple by Prototype-guided Data Synthesis
Junhyeok Choi, Sangwoo Mo, Minwoo Chae
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 PDS instead of optimization-based distillation | TESLA-VL/LoRS · Flickr30K 300 pairs, ResNet | IR@1: 10.3 -> 14.4 | Improved |
| synthesize images with unCLIP decoder from image prototype instead of retrieving real image (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@10: 28.7 -> 37.3 | Improved |
| use unCLIP decoder instead of CLIP-inversion pixel optimization (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@1: 4.4 -> 7.9 | Improved |
| use unCLIP decoder instead of CLIP inversion for generation time (ablation) | PDS · Flickr30K 100 pairs | generation time per image (s): 1477.71 -> 9.71 | Improved |
| filter embeddings to shared pairs when building prototypes (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@10: 28.1 -> 37.3 | Improved |
| add retrieved-caption condition to image prototype (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@10: 28.3 -> 37.3 | Improved |
| discard pairless clusters at larger distilled sizes (ablation) | PDS · Flickr30K 1000 pairs, ResNet | IR@10: 56.4 -> 58.9 | Improved |
| discard pairless clusters at 100 pairs (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@10: 37.3 -> 37.5 | No clear effect |
| use agglomerative clustering instead of mini-batch k-means (ablation) | PDS · Flickr30K 100 pairs, ResNet | IR@10: 37.3 -> 41.4 | Improved |
| replace BERT with CLIP text encoder (ablation) | LoRS baseline · Flickr30K 100 pairs, ResNet | IR@1: 2.0 -> 6.3 | Improved |
| use VAE encoder instead of CLIP encoder (D4M/MGD3 extended) (ablation) | learning-free distillation · Flickr30K 100 pairs, ResNet | IR@10: 37.3 -> 17.2 (MGD3) | Hurt |
| vary guidance scale and sampling steps of decoder (ablation) | PDS · Flickr30K | aggregated retrieval accuracy | No clear effect |
Closest papers by meaning
- Efficient Multimodal Dataset Distillation via Generative Models (NeurIPS 2025 · similarity 0.92)
- CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder (NeurIPS 2025 · similarity 0.91)
- Active Data Curation Effectively Distills Large-Scale Multimodal Models (CVPR 2025 · similarity 0.89)
- Beyond Modality Collapse: Representation Blending for Multimodal Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- CtrlSynth: Controllable Image Text Synthesis for Data-Efficient Multimodal Learning (ICML 2025 · similarity 0.89)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.89)
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.
