ChenhLiwnl

> yeah I think we reversed the `strict=False` and `strict=True`, can you try this? and if it works, would you be willing to open a PR? https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/controlnet_sd3.py#L242 > > ```...

> yeah I think we reversed the `strict=False` and `strict=True`, can you try this? and if it works, would you be willing to open a PR? https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/controlnet_sd3.py#L242 > > ```...

SD3's transformer blocks are: self.transformer_blocks = nn.ModuleList( [ JointTransformerBlock( dim=self.inner_dim, num_attention_heads=self.config.num_attention_heads, attention_head_dim=self.inner_dim, context_pre_only=i == num_layers - 1, ) for i in range(self.config.num_layers) ] ) while sd3_controlnet's transformer blocks are: self.transformer_blocks...

> @wangqixun Can you explain a bit here? The last block is different. In fact the example controlnet (InstantX/SD3-Controlnet-Canny) can be loaded correctly So its a little bit confusing...

sorry but another question, I noticed that in main branch the transformer blocks' attention_head_dim is set to be self.config.attention_head_dim, while in v0.29.2 released version it is self.inner_dim currently self.config.attention_head_dim of...