search

PixArt-alpha / PixArt-sigma

PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation
https://pixart-alpha.github.io/PixArt-sigma-project/
GNU Affero General Public License v3.0
1.44k stars 68 forks source link

why pixart-sigma-256x256 ckpt has no 'pos_embed', while other ckpts has? #37

Closed Gus-Guo closed 1 month ago

Gus-Guo commented 2 months ago

why pixart-sigma-256x256 ckpt has no 'pos_embed', while other ckpts has?

lawrence-cj commented 2 months ago

It must be a bug here. But it would not influence anything since in the ckpt loading during training, we will replace it anyway. https://github.com/PixArt-alpha/PixArt-sigma/blob/ac7004b26f3fce9223ea8fdde1950508c772798c/diffusion/utils/checkpoint.py#L67

Gus-Guo commented 2 months ago

It must be a bug here. But it would not influence anything since in the ckpt loading during training, we will replace it anyway.

https://github.com/PixArt-alpha/PixArt-sigma/blob/ac7004b26f3fce9223ea8fdde1950508c772798c/diffusion/utils/checkpoint.py#L67

I see the code in line 67 replace y_embedder.y_embedding instead of pos_embed, but code in line 57 actually delete pos_embed when loading ckpt. Does it affect inference?

lawrence-cj commented 2 months ago

Nope