Research map / Image generation / Autoregressive and latent generation
Learning from Semantic Dictionaries: Discriminative Codebook Contrastive Learning for Unified Visual Representation and Generation
Imanol G. Estepa, Jesús M. Rodríguez-de-Vera, Bhalaji Nagarajan, Petia Radeva
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 encoder codebook contrast to reconstruction objective (ablation) | LEASE · IN-1K ablation | linear probe accuracy: 73.62 -> 76.11 | Improved |
| add encoder codebook contrast to reconstruction objective (ablation) | LEASE · IN-1K ablation | IS: 79.59 -> 86.71 | Improved |
| apply contrast on decoder features instead of encoder (ablation) | LEASE · IN-1K ablation | FID: 10.62 -> 10.97 | Mixed |
| add decoder contrast on top of encoder contrast (ablation) | LEASE · IN-1K ablation | linear probe accuracy: 76.11 -> 76.07 | No clear effect |
| double codebook size from 8K to 16K centroids (ablation) | LEASE · IN-1K | linear probe accuracy: 75.20 -> 76.70 | Improved |
| use DINOv2 codebook instead of DINOv3 on IN200 (ablation) | LEASE · IN200 (200 ep) | linear probe accuracy: 79.98 -> 85.86 | Improved |
| build DINOv2 codebook on IN200 instead of IN-1K (ablation) | LEASE · IN200 (200 ep) | linear probe accuracy: 85.47 -> 85.86 | Improved |
| raise contrast weight lambda from 0.1 to 1.0 (ablation) | LEASE · IN200 (200 ep) | linear probe accuracy: 85.47 -> 84.93 | Hurt |
| use 15 neighbor centroids instead of 5 (ablation) | LEASE · IN200 (200 ep) | FID: 16.33 -> 16.80 | Hurt |
| use 30 neighbor centroids instead of 5 (ablation) | LEASE · IN200 (200 ep) | linear probe accuracy: 85.47 -> 85.36 | No clear effect |
| LEASE versus MAGE in low-data linear probing | MAGE · IN-1K few-shot | top-1 accuracy: 57.33 -> 59.65 | Improved |
| LEASE versus MAGE under ImageNet-A shift | MAGE · ImageNet-A k-NN | k-NN top-1 accuracy: 4.19 -> 8.51 | Improved |
Closest papers by meaning
- BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities (ICLR 2025 · similarity 0.87)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.86)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.86)
- Leveraging Unpaired Data for Vision-Language Generative Models via Cycle Consistency (ICLR 2024 · similarity 0.86)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.86)
- ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · 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.
