Krea 2 - DC Fantastical Collection SSR Merge
This is a SSR merge from many LoRAs from @dkpc69 , credits to him for training such awesome models.
Here is the list of the LoRAs merged:
DC Aetherion Forge
DC Ancient Futurism Style 1
DC Celestial Dynasty
DC Dark Fantasy Style 1
DC Dark Fantasy Style 2
DC Dark Fantasy Style 3
DC Dark Fantasy Style 5
DC Dark Fantasy Style 6
DC Grim Dark Style 1
DC Obsidian Monochrome
DC Synthetic Anime
DC Vast Expanse
DC Vintage Surreal Fantasy Anime
DC Vivid Dark Fantasy Painting
DC_Fractured_Anime
DC_Liminal_Dreamcore
DC_Retro_Biomech
DC_SCI-FI
When merged, this is a rank 608 LoRA, so this is quite a huge model! There are many, many concepts embedded in this one, so you will need to experiment and take inspiration from the original LoRAs. This model is excellent at merging all the different styles together, acting sort of like a small checkpoint!
Description
Initial version
FAQ
Comments (13)
Awesome work mate did not realise this was possible mixing all these concepts looks good though, Thanks for the mention too
The rank is too high to use this. Krea2 is already a large model. Why not lower the rank to a manageable size, like 1 GB?
The SSR merging technique works like that, there's not much to do about it. It fits on my 12GB GPU, what hardware are you using?
@WhyNaNWhenNyan I use a Mac Mini with 16GB of unified memory. A LoRA of that size would definitely have a severe impact on performance. I know that many people using Krea2 also use the highest quantization levels available to them; consequently, there simply wouldn't be room in memory for that extra 4GB.
If that is how the merging technology works, it might be possible to lower the rank of the already merged LoRA. Alternatively, one could merge it with the base Krea2 model and extract a differential LoRA of various required ranks from it. I would look into this myself, but I work exclusively with quantized base Krea2, and I would suffer a significant loss in accuracy at that stage alone.
@Unicom You can reduce the rank with SVD, but you will lose quality.
The only good solution is to make this merged into a checkpoint.
I've developed this node for comfyui:
https://github.com/frutiemax92/ComfyUI-LoRA-SSR/tree/master
You should be able to use it with your system.
@WhyNaNWhenNyan I dont use comfy at all. Sorry. This thing is extremely inconvenient and not optimized for Apple Silicon. On MacOS / iOS for image generation better to use DrawThings App
@WhyNaNWhenNyan Is it possible to use SSR merge on different models, i.e. Anima?
Yes it is, this works with any set of LoRAs.
@WhyNaNWhenNyan Is there git hub project page or can You share / create tool that You vibe coded?
https://github.com/frutiemax92/LoRAMergingTool
I only did it for Krea 2, you can ask Claude to do the modifications for Anima.
https://github.com/frutiemax92/ComfyUI-LoRA-SSR/tree/master
This node should work with Anima, if it isn't, ping me.