iTARFlow Advances Normalizing Flow Generative Models
A novel generative model known as iterative TARFlow (iTARFlow) has been developed by researchers. This model integrates autoregressive generation with an iterative denoising approach. Unlike traditional diffusion models, iTARFlow employs a fully end-to-end, likelihood-based objective throughout its training phase. For sampling, it utilizes autoregressive generation followed by a denoising process that draws inspiration from diffusion techniques. iTARFlow shows impressive results on ImageNet, achieving competitive performance at resolutions of 64, 128, and 256 pixels. This advancement highlights its capabilities as a robust generative model and pushes the boundaries of normalizing flows. The findings are available in arXiv:2604.20041.
Key facts
- iTARFlow is a new normalizing flow generative model introduced by researchers.
- It combines autoregressive generation with iterative denoising.
- Unlike diffusion models, iTARFlow maintains a fully end-to-end, likelihood-based objective during training.
- During sampling, it performs autoregressive generation followed by iterative denoising.
- iTARFlow achieves competitive performance on ImageNet at resolutions 64, 128, and 256 pixels.
- The model demonstrates potential as a strong generative model.
- The research is published on arXiv with identifier 2604.20041.
- The work advances the frontier of normalizing flows.
Entities
Institutions
- arXiv