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, and is trained as a LoRA natively, so you can apply it to any finetuned model.

It can make Lumina 2 model 2x faster.

Currently support:

NetaYume Lumina (NTYM)

About CFG distilled model:

You can't control CFG scale and negative prompt. Those are inside the model.
No free lunch. There are quality penalties. But less than step distillation (4/8steps model). The most noticeable penalty is that artist styles are weaker. Because artist styles have the highest variance.
It noticeably make hands and limbs better. I assume that distilled model has less "noise and outliners", so it's more stable.
Useful if you want to try new prompt concept fast.

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 for better quality:

Timesteps shift 3 (from node ModelSamplingAuraFlow)
Steps > 20.
Sampler: euler/euler_a + normal, res_multistep + linear_quadratic

FYI, I also write a ComfyUI plugin to enable more hardware accelerations for Lumina 2:

If you have a newer gpu (rtx 30xx and later), you can enable torch.compile to get 30% faster.
If you have an older gpu (rtx 20xx and before) that does not support bf16, ComfyUI by default runs the model in fp32 mode to avoid overflowing, which is extremely slow. This plugin let you run the model in fp16 mode and will take care of any overflows, which is 3~4x faster than fp32.
https://civitai.com/articles/22251

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. Check "Update log" below for more details.

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. In official Lumina 2 code there is a option called "renorm_cfg" and is enabled by default. And I copy-pasted those code. However, actually no one recommends this option and, in fact, leaves it as disabled, because the outcome of this option is questionable. E.g ComfyUI has a node called "RenormCFG" just for Lumina 2, and nobody uses it.
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...

Neta Lumina Lightning LoRA - NTYMv3.5 v1.0