Neta Lumina Lightning LoRA
Details
Download Files
Model description
Neta Lumina Lightning LoRA
All cover images are raw output from the model, 1MP resolution, no upscale, no hands/faces inpainting fixes, even no negative prompt.
What is this
This model is CFG distilled version of Neta Lumina, but trained as a LoRA natively, so you can apply it to any finetuned model. It can make Lumina 2 2x faster.
The cost of CFG distillation is that you can't control negative prompt and CFG scale, they are inside the model. This is acceptable, we don't constantly change negative prompt and CFG scale anyways.
I will try to release multiple sub versions (abcde...), with different CFG (maybe different negative prompt too). So you can choose the CFG you want by choosing LoRA. Check "Update log" below for more details.
If you want to share your images, please post them to the main version page (a).
Currently support:
- NetaYume Lumina (NTYM)
Experimental, but it works.
FYI: This is not a step distillation. Step distillation is too much for me, as well as for the model. Also, step distillation is very aggressive, CFG distillation is simple and has lesser quality penalty.
How to use
Required settings to work:
LoRA strength 1
CFG scale 1 (This disables CFG, and negative prompt)
Do not use cfg++ samplers, they do not work.
Recommended settings for better quality:
Timesteps shift 3 (from node ModelSamplingAuraFlow)
Steps > 20
Tips:
Use this ComfyUI plugin for more optimizations: https://civitai.com/articles/22251
30% faster by using torch.compile.
For older gpu (rtx 20xx and before), 3x faster by running in fp16 mode.
With proper setup Lumina 2 can be "almost" as fast as SDXL. "almost" means 80%.
If you really need to remove something using negative prompt and still want to generate faster. You can split your KSampler to two. First part (20% steps), normal settings with CFG. Second part (rest 80% steps), load this LoRA and set CFG to 1.
License
This model is released under Apache License 2.0.
Additional terms: Selling or monetizing models that merged this LoRA is prohibited.
Update logs
v0.2: Init version
v0.2a: Init version. mimics the effect of CFG 3.5.
v0.2b: This version mimics the effect of CFG 5.
v0.2c: This version mimics CFG 4. Also added some normalizations on mean and variance of noise the teacher model predicted, balabala...
...
Still experimental, can't say which version is better. Try yourself.
You have read this far, don't forget to leave a feedback in discussion. Don't write a review, Civitai review system is so hard to find and, most importantly, read...