Trained using ai-toolkit on vast.ai with an RTX 5090. The process took approximately 12 hours for a total of 20,000 steps.

The ai-toolkit format has been converted for ComfyUI compatibility. Please use the following conversion script: https://huggingface.co/duongve/NetaYume-Lumina-Image-2.0/blob/main/Script_Lora_Convert/Convert_lora_format_between_comfyui_diffusers.py

If you have any questions about the conversion, feel free to ask, or simply consult AI tools like ChatGPT for a quick answer.

This is a project I originally spent four months working on with the SDXL model without completion. Subsequently, due to a heavy workload, there has been no progress for about three or four months. I kept thinking: if I wait until it's perfect to release it, when will it ever be finished?

I did not write longer prompts specifically for Lumina's LLM, so the results might be slightly inferior.

Prompt reference: esophagus, stomach, large intestine, small intestine, rectum, anal sex, large cum, organs x-ray

Perhaps you can switch to others? I mixed in other prompts.

To be honest, regarding structural stability...? It might be better than the SDXL model if the prompts are written to be longer, more structured, and specific.

However, I discovered that training still works well even without the prompt assistant section...?? In the future, I will find ways to synthesize and process higher-resolution images. I am also considering renting an RTX Pro 6000 Blackwell graphics card for high-resolution and larger batch training.

sd-scripts is a bit too difficult to use...

Base model:NetaYume Lumina v3.5