Added ability to load clip pairs randomly from folder. Other small bug fixes
This commit is contained in:
@@ -472,8 +472,8 @@ class VisionDirectAdapter(torch.nn.Module):
|
||||
|
||||
self.mid_size = self.token_size
|
||||
|
||||
# if pixtral, use cross attn dim for more sparse representation
|
||||
if is_pixtral:
|
||||
# if pixtral, use cross attn dim for more sparse representation if only doing double transformers
|
||||
if is_pixtral and self.config.flux_only_double:
|
||||
if is_flux:
|
||||
hidden_size = 3072
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user