The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Use empty t5 of size 128 for hidream, seems to give closer results.
comfyanonymousfd274944418f
diff --git a/comfy/text_encoders/hidream.py b/comfy/text_encoders/hidream.pyindex ca54eaa78..8e1abcfc1 100644--- a/comfy/text_encoders/hidream.py+++ b/comfy/text_encoders/hidream.py@@ -120,7 +120,7 @@ class HiDreamTEModel(torch.nn.Module):ll_out = Noneif t5_out is None:- t5_out = torch.zeros((1, 1, 4096), device=comfy.model_management.intermediate_device())+ t5_out = torch.zeros((1, 128, 4096), device=comfy.model_management.intermediate_device())if ll_out is None:ll_out = torch.zeros((1, 32, 1, 4096), device=comfy.model_management.intermediate_device())