Skip to content

AttentionProcessor.group_norm num_channels should be query_dim #3046

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Apr 11, 2023
Merged

AttentionProcessor.group_norm num_channels should be query_dim #3046

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
072a5dd
`AttentionProcessor.group_norm` num_channels should be `query_dim`
williamberman Apr 10, 2023
33b4c02
add_{k,v}_proj should be projecting to inner_dim
williamberman Apr 10, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions src/diffusers/models/attention_processor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,7 @@ def __init__(
self.added_kv_proj_dim = added_kv_proj_dim

if norm_num_groups is not None:
self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
else:
self.group_norm = None

Expand All @@ -93,8 +93,8 @@ def __init__(
self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)

if self.added_kv_proj_dim is not None:
self.add_k_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
self.add_k_proj = nn.Linear(added_kv_proj_dim, inner_dim)
self.add_v_proj = nn.Linear(added_kv_proj_dim, inner_dim)

self.to_out = nn.ModuleList([])
self.to_out.append(nn.Linear(inner_dim, query_dim, bias=out_bias))
Expand Down