NOTE - Final Version of the 4-Step LoRA was released today (folders and files restructured, renamed and cleaned up)
The main LoRA files no longer contain chkNNN or _latest in the name, now the names are:
krea2_turbo_4step_rank_64_lora_comfyui.safetensors
krea2_turbo_4step_rank_64_lora.safetensors
checkpoint_resolution_sweeps folder is now renamed and moved under assets/resolution_sweeps
older_checkpoints folder is now moved under _archive where you can see both the checkpoints and their respective resolution sweeps - demo images.
The README.md has been completely rewritten with this being final version in mind.
Licensing information has been added (to be compliant with Krea's own licenses).
Other than that, the final LoRA achieves the following - Half the steps 路 ~1.6脳 faster 路 45% of the gap to the 8-step teacher closed 路 texture at 1.03脳 the teacher's, verified clean 路 in VLM tests the teacher preferred on only 6 of 45 judged renders 路 12 trained resolutions 路 78,000 training samples on 13,750 teacher trajectories and 43,044 real-photo crops 路 21 days on one RTX 3090. Yes a mouthful 馃槂
It has been an amazing project to work on, and I hope you all enjoy the result as much as I do.
// Note - I am giving it my best on the next (try) project - 2 step LoRA of the same (it is unlikely it would achieve the same quality as this 4 step LoRA, but I will try my best to get it to generate good enough images for general fast 2 step previews). No guarantees it would work! :)
In the meantime have fun with the 4 Step LoRA!
With 4 steps, it's already so good, awesome to generate 1MP image in under 4 secs on an overclocked RTX 3090.