The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
fix a bug in the attn_masked redux code when using weight=1.0 (#6721)
Raphael Walkerb6951768c41f
diff --git a/nodes.py b/nodes.pyindex ba9c4e4bb..9779d5fdb 100644--- a/nodes.py+++ b/nodes.py@@ -1064,7 +1064,8 @@ class StyleModelApply:for t in conditioning:(txt, keys) = tkeys = keys.copy()- if strength_type == "attn_bias" and strength != 1.0:+ # even if the strength is 1.0 (i.e, no change), if there's already a mask, we have to add to it+ if strength_type == "attn_bias" and strength != 1.0 and "attention_mask" not in keys:# math.log raises an error if the argument is zero# torch.log returns -inf, which is what we wantattn_bias = torch.log(torch.Tensor([strength]))