Uh oh!
There was an error while loading. Please reload this page.
Improve control net block index for sd3 - #9758
Conversation
linjiapro
commented
Oct 23, 2024
linjiapro
commented
Oct 23, 2024
This is a very simple PR, but for some reason, the tests all failed for wired reasons such as: Unable to find self-hosted runner group: 'aws-general-8-plus'. |
sayakpaul
commented
Oct 24, 2024
Thanks for your contributions! Could you maybe also add a test for this? |
HuggingFaceDocBuilderDev
commented
Oct 24, 2024
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
linjiapro
commented
Oct 24, 2024
There is an existing test for the pipeline of SD3 control net, I leveraged that. The layer number for the control net changed from 1 to 3, the number of layer of the transformer (4) is no longer the multiples of the number of layers of control net. This will test the code changes of this PR. |
sayakpaul
left a comment
There was a problem hiding this comment.
Thanks, I left some comments.
| ENV/ | ||
| env.bak/ | ||
| venv.bak/ | ||
| myenv/ |
| patch_size=1, | ||
| in_channels=8, | ||
| num_layers=1, | ||
| num_layers=3, |
There was a problem hiding this comment.
We should change this value here I think. Instead, we could make this method accept an argument like num_controlnet_layers and then leverage it as needed. WDYT?
@sayakpaul@yiyixuxu, can we take a look at this? Thanks |
bghira
commented
Nov 16, 2024
maybe @DN6 is more active. |
| # controlnet residual | ||
| if block_controlnet_hidden_states is not None and block.context_pre_only is False: | ||
| interval_control = len(self.transformer_blocks) // len(block_controlnet_hidden_states) | ||
| interval_control = len(self.transformer_blocks) / len(block_controlnet_hidden_states) |
There was a problem hiding this comment.
why are we making this change? it is not the same so a breaking change, no?
There was a problem hiding this comment.
@yiyixuxu Good question.
The revised code adapts the strategy used by ControlNet for Flux, introducing a significant improvement in flexibility. Here's why this change matters:
In the old code, the number of transformer layers is divisible by the number of ControlNet layers. For example, with SD3.5 Large, which has 38 transformer layers, there were only two valid options for the number of ControlNet layers: 2 and 19. Setting the number of ControlNet layers to anything else, such as 5, would cause the old code to crash.
However, the Flux ControlNet approach removes this restriction, allowing greater flexibility in choosing the number of layers. The revised logic essentially mirrors the Flux implementation, enabling more versatile configurations.
Importantly, the new code maintains compatibility with existing setups. If the number of transformer layers is divisible by the number of ControlNet layers, the interval_control remains unchanged, ensuring all previous configurations continue to function seamlessly.
There was a problem hiding this comment.
thanks! I think it's indeed better, I'm just wondering if it would cause issue for controlnet is trained with the current logic
cc @haofanwang here
There was a problem hiding this comment.
I don't think it will cause any issue with the trained controlnet using old code before this PR.
The reason is that for the controlnet to be trained with the old code, the number of layers of the transformer has to be divisible by the number of layers of the controlnet, and the new logic after this PR does not change the behavior for the above scenario.
Uh oh!
There was an error while loading. Please reload this page.
Co-authored-by: YiYi Xu <yixu310@gmail.com>
yiyixuxu
commented
Nov 20, 2024
can you run |
* improve control net index --------- Co-authored-by: YiYi Xu <yixu310@gmail.com>
What does this PR do?
The layer configuration for the Control Net in Stable Diffusion 3 models must adhere to the rule that the total number of layers in the SD3 model should be a multiple of the Control Net's layer count.
For SD3.5, which has 38 layers, the Control Net can only have three possible options: 2, 19, or 38. This leads to inefficiencies in the setup.
Also, qk_norm, context_pre_only_last_layer, dual_attention_layers are added to match the transformer architecture
Who can review?
@sayakpaul@yiyixuxu@DN6