Skip to content

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 black images on some flux models in fp16. (#8495)

comfyanonymousaf1eb58be8ce
diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py
index 53f27e3a7..09dd2482c 100644
--- a/comfy/ldm/flux/model.py
+++ b/comfy/ldm/flux/model.py
@@ -187,6 +187,9 @@ class Flux(nn.Module):
if add is not None:
img[:, txt.shape[1] :, ...] += add
+ if img.dtype == torch.float16:
+ img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
+
img = img[:, txt.shape[1] :, ...]
img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)