Regarding artist mixing through the cross-attention layer of hook Anima
Anima has caught my attention since its preview 1 release. It is an open-source anime model with a brand-new architecture, but its artist mixing capability is somewhat lacking. Over the past few days, I have pondered many solutions. Ultimately, by hooking into Anima's cross-attention layer, I have achieved controllable mixing of conditioning from multiple artists during the cross-attention stage, solving the problem of LLM TE struggling to correctly understand artist combinations when directly stacking multiple artist tags in the prompt. After testing with my collaborative developers, the results seem promising. I will further optimize this plugin and officially release it in the next few days!（By the way, I haven't used GitHub yet and might need to learn it first, as this is my first attempt to develop a plugin with the assistance of AI）
Very interesting! I’m also working on a similar custom node. It seems to work to some extent, but when the number of tags increases — around 10 or more — they don’t blend well and the result starts to look flat. Have you experienced a similar issue? https://github.com/toyxyz/ComfyUI_toyxyz_test_nodes
Dose your node has github link?
Dose your node has github link?
I am still continuing to optimize this node. Once the optimization is completed, I will upload it to GitHub
Very interesting! I’m also working on a similar custom node. It seems to work to some extent, but when the number of tags increases — around 10 or more — they don’t blend well and the result starts to look flat. Have you experienced a similar issue? https://github.com/toyxyz/ComfyUI_toyxyz_test_nodes
I haven't tested with so many artist tags yet, but the fundamental reason should still be a mathematical issue. Anima cannot achieve the kind of artist mixing like SDXL
Very interesting! I’m also working on a similar custom node. It seems to work to some extent, but when the number of tags increases — around 10 or more — they don’t blend well and the result starts to look flat. Have you experienced a similar issue? https://github.com/toyxyz/ComfyUI_toyxyz_test_nodes
I haven't tested with so many artist tags yet, but the fundamental reason should still be a mathematical issue. Anima cannot achieve the kind of artist mixing like SDXL
By the way, it seems that directly intervening in the text encoder part is not feasible. I have been trying this approach before intervening in the attention layer