Krea 2 - YARN Collection SSR Merge
This is a merge of all the excellent LoRAs from @TimmyHodor Krea 2 YARN Collection. This merges the 17 LoRAs all together, using the Subspace Signal Routing technique from this paper, vibecoded for Krea 2 with Claude:
https://arxiv.org/abs/2606.10617
This creates a huge LoRA with rank of 32*17=544! This is not a simple averaging merge, it finds weights that are specifically important for some LoRAs, so it avoids in-between models corruption.
When you combine a bunch of awesome models, you get a really awesome, multi-purpose model!
Description
17 merged models from the YARN collection from @TimmyHodor
FAQ
Comments (9)
This is the first I've seen of this paper. Please do more. I'm super curious how versatile it is.
I will, I cannot train proper Krea 2 models on my hardware, but SSR works fine! I'm really impressed so far by the results.
It's like thanos finally snaps that finger
Great results, I love this :-).
THX, Your the Best
So the original loras don't include any trigger words. How would I prompt to trigger the right concept/style with this lora? Or am I just gonna get random concepts?
This should be used as a general purpose NSFW model. Some outfits have been preserved such as the BLACKED and Zishy bras, etc.
You won't get the stylized lighting from the BLACKED LoRA for example, but you will get an average of all models. Feel free to experiment with this model by using the prompts from the original YARN collection LoRAs.
How is this different from https://civitai.red/models/2880467/the-yarn-collection-for-krea2?
This is a merge of all of these LoRAs, using the SSR merge technique.