IMPORTANT: For ComfyUI, you MUST download the optional Comfy lora instead (see optional downloads above).
This LoRA is the extracted difference between the base and turbo Krea 2 at dim 256.
Mostly it's intended to use as an inference lora with my fork of ai-toolkit:
https://github.com/envy-ai/ai-toolkit-envy-optimized
Feel free to upload this wherever or mess with it or do whatever you want, as long as you're complying with Krea 2's model license.
Description
FAQ
Comments (11)
A lot of turbo LoRAs actually work even in very low ranks. Worth trying. Might depend on type of "turbo" used.
Apparently there's a dim64 one floating around that I didn't know about before I made this. :)
Here is the dim64 version: https://huggingface.co/Comfy-Org/Krea-2/blob/main/loras/krea2_turbo_lora_rank_64_bf16.safetensors
I wonder what are the benefits of using the Turbo LoRA on the raw checkpoint compared to the Turbo checkpoint 🤔
Just to be clear, there are two LoRAs that can be downloaded:
1. krea2_raw_to_turbo_r256.safetensors: this is to be used with AIToolkit.
2. Under "optional files", krea2_raw_to_turbo_r256_comfy.safetensors: this is to be used with ComfyUI.
Make sure you download the right one.
"it's intended to use as an inference lora with my fork of ai-toolkit:"
so no lora training use?
@wktra My guess is that by using the LoRA, one can train using RAW, but generate the epoch samples using RAW + this turbo LoRA. That way the trainer can see how the LoRA will look like when used with the Turbo model (most people generate using turbo rather than raw).
One thing I noticed that while the Krea 2 Raw model is limited generates up to 1K resolution(unlike it's turbo varient that generates up to 2K), I find that using the Turbo LoRA with it allows me to generate up to 2K resolution without getting artifacts. Is this magic?
Here are the recommended settings I use with the Turbo LoRA on the Raw Krea 2 checkpoint: 0.6 weight, and 12 steps at 1CFG. This allows for a bit more diversity and adherence imo.