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 is CFG distilled Neta Lumina, and is trained as LoRA natively.
2x faster.
Currently support:
NetaYume Lumina (NTYM)
About CFG distilled model:
You can't control CFG scale and negative prompt. Those are trained inside the model.
Built-in artist styles are weaker. Because artist styles have the highest variance.
Other kinds of penalty (concept etc.) are unnoticeable. Unlike step distillation (4/8 steps model etc.), CFG distillation is simple and not aggressive.
Interesting pros:
Noticeably better hands and limbs.
Stronger and more accurate (?) style from style LoRA.
Don't know why. I assume that distilled model has less "noise and outliers", so overall it's more "stable".
How to use
Required settings to work:
LoRA strength 1
CFG scale 1 (This disables CFG, and negative prompt)
Do not use cfg++ (cfg_pp) samplers, they do not work.
Recommended settings:
Timesteps shift 3 (from node ModelSamplingAuraFlow)
Steps 20. If you really just want to test something (concept etc.) very quickly, 12 steps is also doable.
Sampler: euler/euler_a + normal
License
This model is released under Apache License 2.0.
Additional terms: Selling or monetizing models that merged this LoRA is prohibited.
Update logs
I may release multiple sub versions (abcde...), with different CFG (maybe different negative prompt too). So you can choose the CFG you want by choosing LoRA.
If you want to share your images, please post them to the main version page (a).
v1.0 (11/21/2025):
Mimics the effect of CFG 5.
Finally, found and fixed a noise mismatch issue in all v0.2 versions. What the f is the "renorm_cfg" thing in official Lumina 2 code.
Also skipped early layers which have extremely large activation values. Should have better compatibility with other LoRAs.
v0.2: Init version, experimental.
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...
Description
FAQ
Comments (13)
Very nice work. This brought my average gen time down from 39s to 21s - not quite cut in half, but very close! I tested it on Neta Cat Tower as well as Neta Yume 3.5 & it worked fine.
There's definitely a slight quality hit - I got a muddy look on some images, and maybe a slightly increased chance of anatomy failure (though it's hard to tell if that's due to the lora or just bad luck), but this is a great start.
Settings I used were 30step/cfg1/res_multistep/linear_quadratic
I think "res_multistep" sampler is not supported. Please try eular or eular_a.
@reakaakasky I tested Euler and Euler Ancestral as well and didn't see any real improvement. res_multistep did fine & overall was better quality.
@gothefungus872 I see. res_multistep looks fine with linear_quadratic, but not fine with the "normal" scheduler.
This is strange. I tested it with "normal" scheduler, there were lot of artifacts so I thought it didn't work.
Anyways, euler is always recommended because it's guaranteed to work. I will update the page.
Btw, dpm2 / dpmpp_sde / seeds3 works even better, but they are 2-3 times slower then euler/res_multistep.
"a slightly increased chance of anatomy failure"
v0.2b may have fixed this
This may be a game changer for this model viability =0
which version is better?
hi, some folks on neta and furry diffusion discord have tested the loras, and there's definite - even semi drastic quality drop compare to your claim of non dropoff in performance. How did you achieved it? Also, do you have a discord =D , suprised to see you be very involved in neta but I cant find you anywhere beside here.
@RicemanT no. I'm not active on discord.
> claim of non drop in performance
I admit that statement was too exaggerate. That was based on a preliminary test, and the prompts were very simple. I've done more tests since then and found quite a few problems too, such as style shifting.
In my tests v0.2b seems to be better than v0.2.
v0.2 has more artifacts on small details like fingers.
@reakaakasky ok, good to know regardless. Another guy that tested it recently claim the quality drop with your lora got more managable when combined with another lora (?). Nevertheless, appreciate the work very much!
v0.2b seems to be better than v0.2c.
v0.2c seems to have even more artifacts than v0.2.
v0.2b is very stable. I'd say v0.2b gives me the same quality as the original model, if I don't care about the style. With teacache, I can reduce model runtime by 3/4. It's perfect for quickly trying out new prompt concepts.
Appreciate the work very much!