Enhancement bundle for Lumina

Hello Lumina Image 2.0. Goodbye SDXL.

I know only gigachad users would see and read this description for now. So I gonna write this in a straight way.

I may rewrite this page later.

Lumina Image 2.0 is a amazing open-source base model released by the Alpha-VLLM team, on Feb, 2025.
Neta Lumina is anime-style finetuned model from Lumina Image 2.0.
NetaYume Lumina is finetuned from Neta Lumina.

If you just heard about Lumina Image 2.0 and wonder the differences between it and SDXL:

Newer, released in Feb 2025. (vs July 2023)
Better VAE from Flux. (Mathematically 4x details)
Better model arch. Slightly larger (1.25x) model size, but much stronger performance. (5GiB vs 4GiB, bf16)
Slower than SDXL. Needs 3x computational time.
Quantized models work very well. No CNN layers. So you can run a Q8 model, which is only 3GiB, without noticeable downgrade in quality. Even Q4, if you want.
Much much better prompt understanding. (Gemma-2-2B vs CLIP-L/g, and new arch)

In short: Except the 3x computational time, Lumina 2 is just way better. After all, SDXL is just too old, and it's arch is not efficient as new arch.

Same as my previous "stabilizer" model. (But with a less misleading name. lol)

Finetuned from NetaYume Lumina v3. But as a LoRA.
Trained with mixed dataset, which includes a little bit everything.
Try to alleviate the problem that anime model was biased to over simplified anime data.
Does not have default style (no default bias). You need to add style trigger words (artist name etc.)

Effect:

It should gives you better backgrounds, natural textures, better hands, less noise and maybe less deformed images. Still testing.
But don't expect it would make huge difference like my stabilizer LoRA. The lumina 2 base model does not have big overfitted problem like illustrious/NoobAI.

For those who want to give it a try:

Tips of using Lumina 2, from my exp:

Unlike SDXL, it requires a very detailed prompt. Simple prompt gives deformed images. I use a LLM to automatically enhance my prompt.

Training:

This LoRA was trained using kohya-ss/sd-scripts.

Lumina training script is not officially release yet, right now (9/21/2025) the script is under sd3 branch. You can use it out-of-the-box.
If you want to train a LoRA as well. Just follow the official document.
It works, but outside the official document there still has some small problems, some settings are not working and buggy. (need time to fix)
And some improvements can be made, e.g. +50% training speed. (I've already implemented and tested it, ready to pull)
I did not add system prompt. I don't think it's good idea, because original Lumina Image 2.0 trained without it.

I'm still testing. For now I'm only using ~1k images, not my full dataset.

This model is released under Apache License 2.0.

Additional terms: Selling or monetizing models that merged this LoRA is prohibited.

Description