Enhancement bundle for Lumina 2

All cover images are raw output from the model, 1MP resolution, no upscale, no hands/faces inpainting fixes, even no negative prompt.

Currently support:

• NetaYume Lumina (NTYM)

Finetuned with ~7k images.

• Many specialized sub datasets, such as close-up clothing, hands, complex ambient lighting ...

• Only high resolution images with finest details. Every image is hand-picked by me.

• Comprehensive natural language captions from Google LLM. Anime characters are tagged by wd tagger v3 first and then refined to natural language by Google LLM.

Effects:

• Better backgrounds, natural textures, lighting, and less noise.

• Maybe slightly better creativity and prompt following.

• Standard noise schedule. Inpainting/img2img friendly.

For now there are two kinds of version:

• No-bias: The dataset is very diverse. By default the finetuned model does not have default style (aka. bias). And is very "creative". You need to specify styles in your prompt (artist, style tags etc. or style LoRAs) to guide the model and usually you can get better result (better prompt adherence, more accurate style, better details etc.). E.g. Jojo x frieren https://civarchive.com/images/108073786 .

• Has default anime style: Anime images have higher repeats in the dataset. So there is a default anime style (bias). Easy to use, but less "creative", because the model will always try to generate the default anime style no matter what you prompt. Useful if you just want a general anime style and don't want to prompt style.

How to use

Strength 1 is recommended.

For anime image in the dataset, I used "Digital anime illustration." in prefix. You can use this as trigger words.

Other tools:

• Lightning LoRA for Neta Lumina 2, 2x faster.

• ComfyUI optimization plugin for Lumina 2: 3x faster on older gpu (rtx 2xxx and before), 30% faster on newer gpu (rtx 3xxx and later).

License 

This model is released under Apache License 2.0.

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

Update logs

v0.37: No bias version.

v0.27: Has a default anime style. Stronger overall effects than v0.11.

v0.11: Trained on NetaYume v3.5. Full run. Trained with all my dataset. Has a default anime style.

v0.8: init version. Trained on NetaYume v3. Testing run. Total training data ~1k. No bias, no default style.