New adaptive method for pixel diffusion transformers with MFlow
In brief
- MFlow adaptively combines x0- and u-prediction based on data principal directions.
- ConvRefiner post-hoc module corrects high-frequency defects in generated images.
- Competitive results on ImageNet, UCF-101, and text-to-image generation.
The authors introduce MFlow, an adaptive parameterization for pixel-space diffusion transformers that blends x0- and u-prediction based on data directions. They also propose ConvRefiner to fix high-frequency deficits in generated images. Results show competitive performance across ImageNet, UCF-101, and text-to-image generation.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error