Uncensored KREA2
Examples images shows the base model (int8) on the left / finetune (int8) on the right (same seed same prompt, 8 steps)
This V3 was retrained, meant to be semi-realistic, have less aesthetic than V2, but finer details, was trained at all available buckets from 512 to 2048 px, with a different methodology than V2.
Then was blocks filtered, tweaked with a tool I have made that refine blocks weights by manual iterations over multiple rounds of A/B comparison discrimination.
You can use my "Magic LoRA Loader" from https://github.com/PGCRT/CRT-Nodes
If you want inference blocks tweaks
Tip - For sharper output, you can use the Wan2.1_VAE_upscale2x_imageonly_real_v1 with https://github.com/spacepxl/ComfyUI-VAE-Utils Loader / decoder
Get free credit between 5-500$ in runpod with my link https://runpod.io?ref=u7b2habt
Or buy me a coffee If you want to support my work. Have fun
Description
This V2 was retrained, Extracted LoRA with filtered blocks.
FAQ
Comments (12)
Thanks for your work, would you please share some prompts so we can align?
I'm also somewhat lost as to what this Lora does. Nice images though
https://drive.google.com/file/d/1hE11GF4esPTlkXi6ZifRPiVucbXqHmGd/view?usp=sharing
These prompts are generated by gemma4 26b abliterated with instruction shift at every run, and I didn't even check 5% of it, so good luck
I included the workflow used to generate these example images, add the path of the prompts folder and let it run
@pgc Thanks, just curious!
This is a lora right. It's just posted as a checkpoint, but used like a lora, right
Looks like V2 is the Lora extracted from his main BF16 model, form what i understand.
really love your work keep going man <3
nice model. int8 plz.
Holy shit!! really a sauce material.
How is this only 300 MBs, what am I missing?