1girl, solo, blue archive, long hair, pink hair, purple hair, light blue hair, multicolored hair, two-tone hair, streaked hair, twintails, drill hair, blue eyes, purple eyes, ahoge, halo, wing hair ornament, star hair ornament, dress, bow, star (symbol), puffy short sleeves, jacket, purple bow, white gloves, striped pantyhose,
purple shoes, wing footwear ornament
11.11 : vpred v1.1 update. Increased training dataset to 123 images, applied LoKr algorithm.
The advantage of LoKr is its compression capability, derived from high parameter efficiency. This gives it the potential to achieve equivalent or even greater expressive power than standard LoRA, despite having a file size that is approximately 8 times smaller.
10.30 : trained with 96 images
This is an experiment to compare the differences between the IL and Vpred LoRAs.
All training parameters for both LoRAs are identical, except for the difference in the learning models (IL 0.1 and NoobAI v-pred 1.0).
Overall, the IL side exhibits bright colors and thin lines (this look is somewhat closer to the 'Blue Archive' style), while the vpred side shows deep shadows/contrasts and thick lines.
If this washed-out/overexposed appearance (turning white) is due to the white background in the dataset, the vpred side could be considered to have learned the character features more effectively. However, since the IL model is also not bad for general use, it is difficult to determine which one is superior.
