RAPHAEL
“
When one is painting one does not think.
”
—
Raffaello Sanzio da Urbino
https://raphael-painter.github.io/
Project
https://arxiv.org/abs/2305.18295
RAPHAEL: Text-to-Image Generation via Large Mixture of Diffusion Paths
https://gyazo.com/ac6da6d4a5ca0d087fbfcfbd262d67f9
4つのコンポーネントからなる
Self-Attention
Cross-Attention
Time-MoE
Space-MoE
(図3:(b))
COCO dataset
で学習
比較
Stable Diffusion XL
/
DeepFloyd IF
/
DALL-E 2
/
ERNIE-ViLG
https://gyazo.com/6d6a5a296dbbf7732bd3f7cfaa1f5ab2
text2image