Research map / Image generation / Autoregressive and latent generation
Benchmarking and Analyzing Generative Data for Visual Recognition
Bo Li, Haotian Liu, Liangyu Chen, Yong Jae Lee, Chunyuan Li, Ziwei Liu
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 defined (dataset-specific) templates instead of simple template (ablation) | Stable Diffusion · GenBench 22 datasets mean | CLER score (%): 48.83 -> 51.60 | Improved |
| use defined templates instead of simple template (ablation) | GLIDE · GenBench 22 datasets mean | CLER score (%): 48.48 -> 52.04 | Improved |
| use category enhancement (T5 sentence expansion) instead of simple template (ablation) | Stable Diffusion · GenBench 22 datasets mean | CLER score (%): 48.83 -> 47.68 | Hurt |
| use category enhancement instead of simple template (ablation) | GLIDE · GenBench 22 datasets mean | CLER score (%): 48.48 -> 46.93 | Hurt |
| add negative prompts on top of defined templates (ablation) | Stable Diffusion · GenBench 22 datasets mean | CLER score (%): 51.60 -> 51.10 | No clear effect |
| add negative prompts and restrictive descriptions on top of defined templates (ablation) | Stable Diffusion · GenBench 22 datasets mean | CLER score (%): 51.60 -> 52.01 | No clear effect |
| add negative prompts and restrictive descriptions (ablation) | Stable Diffusion · ImageNet-1K and CIFAR-10 (20-shot) | CLER score (%): 61.07 -> 61.45 | Improved |
| use Stable Diffusion 2.1 instead of GLIDE with simple prompt | generative data · GenBench 22 datasets mean | CLER score (%): 48.48 -> 48.83 | No clear effect |
| inject original data (5-shot) via Textual Inversion | Stable Diffusion · GenBench 17 datasets mean (5-shot generated) | CLER score (%): 45.92 -> 50.52 | Improved |
| inject original data (5-shot) via Textual Inversion | Stable Diffusion · GenBench 17 datasets mean (100-shot generated) | CLER score (%): 48.64 -> 53.90 | Improved |
| inject retrieval data via Textual Inversion | Stable Diffusion · GenBench 17 datasets mean (5-shot generated) | CLER score (%): 45.92 -> 46.46 | Improved |
| inject retrieval data via Textual Inversion | Stable Diffusion · GenBench 17 datasets mean (100-shot generated) | CLER score (%): 48.64 -> 47.75 | Hurt |
Closest papers by meaning
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.90)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.89)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.88)
- LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models (NeurIPS 2026 · similarity 0.88)
- GenAI Arena: An Open Evaluation Platform for Generative Models (NeurIPS 2024 · similarity 0.88)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.87)
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.
