Researchers have investigated the coupling between tokenizers and generators in medical image generation, specifically on the ChestMNIST dataset. Their findings indicate that the optimal tokenizer choice is dependent on the specific generator and sampler used, challenging the assumption of a fixed tokenizer. They introduced a new metric, neighbor-conditional predictive gain, which effectively distinguishes tokenizer families based on downstream generation quality, outperforming reconstruction PSNR and marginal token entropy. Experiments on LFQ-1024 with D3PM and SE-D3PM models showed significant improvements in FID scores when retuned with sampler selection. AI
IMPACT This research could lead to more accurate and efficient medical image generation models by optimizing the interaction between different model components.
RANK_REASON The cluster contains a research paper detailing novel methods and findings in medical image generation. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- ChestMNIST
- D3PM
- DagsHub
- Gotit.pub
- Hugging Face
- LFQ-1024
- ScienceCast
- SE-D3PM
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
