Research map / Label-efficient and robust learning / Recognition and long-tail
DiT-Distill: Open-Set Fine-Grained Retrieval via Generative Curriculum Knowledge
Xin Jiang, Hao Tang, Meiqi Cao, Junyao Gao, Fei Shen, Zechao Li
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 Generative Curriculum Distillation of DiT knowledge to ViT-B/16 student (ablation) | ViT-B/16 · CUB-200-2011 | Recall@1: 77.4 -> 87.2 | Improved |
| add Generative Curriculum Distillation of DiT knowledge to ViT-B/16 student (ablation) | ViT-B/16 · Stanford Cars | Recall@1: 72.8 -> 91.4 | Improved |
| add generative infusion module with vanilla DiT teacher at inference (ablation) | ViT-B/16 · CUB-200-2011 | Recall@1: 77.4 -> 85.3 (time 2.6ms -> 33.5ms) | Improved |
| add conditional discrepancy refinement (CDR) to the DiT teacher (ablation) | Teacher-assisted model · CUB-200-2011 | Recall@1: 85.3 -> 86.5 | Improved |
| add curriculum alignment loss to distill and discard DiT (ablation) | Refined teacher-assisted model · CUB-200-2011 | Recall@1: 86.5 -> 87.2 (time 33.5ms -> 2.6ms) | Improved |
| use 32 learnable queries instead of 16 or 64 (ablation) | Generative Infusion Module · CUB-200-2011 | Recall@1: k=4 86.5, k=16 86.9, k=32 87.2, k=64 87.0 | Improved |
| use 3-layer bridging encoder instead of MLP (n=0) (ablation) | Generative Infusion Module · CUB-200-2011 | Recall@1: 85.1 -> 87.2 | Improved |
| increase bridging encoder depth from 3 to 6 layers (ablation) | Generative Infusion Module · CUB-200-2011 | Recall@1: 87.2 -> 86.7 | Hurt |
| add coarser timesteps to the curriculum (t4 only to t2-t4) (ablation) | Curriculum alignment loss · CUB-200-2011 | Recall@1: 86.7 -> 87.2 | Improved |
| add noisiest timestep t1 to the curriculum (ablation) | Curriculum alignment loss · CUB-200-2011 | Recall@1: 87.2 -> 87.2 | No clear effect |
| use manual bounding boxes instead of detector for CDR data (ablation) | DiT-Distill · CUB-200-2011 | Recall@1: 87.2 -> 87.4 | No clear effect |
| set distillation weight alpha too low or too high versus 2.0 (ablation) | DiT-Distill · CUB-200-2011 | Recall@1 | Hurt |
Closest papers by meaning
- See Further When Clear: Curriculum Consistency Model (CVPR 2025 · similarity 0.85)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.84)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.84)
- Fine-Grained Distillation for Long Document Retrieval (AAAI 2024 · similarity 0.84)
- Generative Multi-Modal Knowledge Retrieval with Large Language Models (AAAI 2024 · similarity 0.84)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.84)
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.
