Home trained FLUX, SDXL and Z-IMAGE Turbo Lora on a RTX 5070 ti graphics card.
Online trained KREA2.
SDXL version 1.0: Trained with Image Dataset (using OneTrainer)
FLUX version 1.0: Trained with Image Dataset (using FluxGym)
FLUX version 2.0: Fine-tuned Lora more refined oriental illustrations
FLUX version 3.0: Fine-tuned Lora to add Illuminated manuscript borders and more refined oriental illustrations
ZIT version 1.0: Trained with Image Dataset (using musubi-tuner)
KREA2 version 1.0: Trained with Image Dataset (using CIVITAI) Style is less consistent on merged base models trained on realistic photography, recommended to use Krea2 Turbo FP8 model on long narrative prompting. Otherwise Keep prompting short text with enough trigger tags on the merged base models.
Description
SDXL version 1.0:
1 repeates, 100 epochs, 58 images, batch 2 = 2900 steps (Size Images Dataset 1024x1024)
Training images are created with ComfyUI using specific prompts using the Flux version of this LORA and slightly mixed with Memadukan Brush Strokes of Fineart and Satria Gandrong. It's generates landscapes and buildings, but does not generate people very well at this point; as brush strokes might deform faces. This is still work in progress for future fine-tuning.
Showcase (SDXL) are results from the SDXL LORA itself.