Qwen Lightning 8 steps
https://huggingface.co/lightx2v/Qwen-Image-Lightning
Usage:
Strength: 1.0
CFG: 1
Steps: 4 or 8
Sigma Shift: 0.3-3YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Description
FAQ
Comments (13)
Hi will this work for GGUF 3Q KM and Q5 KM checkpoints?
also works with distilled FP8 model. Chaining distilled and lightning together enables you to go even down to 6 steps
Mind posting your workflow file? I'm getting terrible results using "all the same settings".
FIGURED IT OUT. If you're getting godawful results at 8 steps, check your terminal outputs. If you see "lora key not loaded" at the start of your render, you need to update your comfyui (I switched to nightly build, but may be fixed on main now, don't know). Don't know how this will affect swarm as I'm just using naked comfyui + comfy manager with qwen for now.
With that fix in place, getting BEAUTIFUL images at 8 steps.
Best LoRA ever! 10+ times faster than the Qwen base model, with no significant loss in quality! It even runs on my Mac Mini superfast using the Apple Silicon, thanks to this amazing python tool from Ivan Fioravanti (a 1 liner to setup all and generate the image using the Lightning LoRA!) https://github.com/ivanfioravanti/qwen-image-mps
DUDE where is the workflow?
Works fine👍🏻
CFG 1?
Works well. Using a distilled model with the 4-Step Lora you can generate an image with as little as only 2 steps. With the Q3_K_M GGUF a 2MP image is 25 seconds on RTX 4070 TI 12GB VRAM.👍
Anyone also gets "lora key not loaded" and not working?
right now i can use this lora without the annoying error but the lora doesn't really speed up the generation time for some reason. does anyone have any success with this lora ?
I am getting blurry images like this LORA do not work;/ Any ideas how to fix it?
Thanks you very much darksidewalker for your very great job with this lora. Time went from 40 minutes per image to 5 to 7 minutes. (GtX 1050 TI 4GB VRAM) . You are the best, man :)
* 8steps v.1.0 here and on huggingface are different (size and hash), is there some reason?