This LoRA is designed to generate handwritten Japanese text in speech bubbles, taking advantage of Krea2's Japanese text processing capabilities.For v0.0.0, I trained the LoRA on 60 speech bubbles featuring handwritten text by <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/artists/94255">Hanasaka Houcha</a>. The model has successfully learned some of the characteristics of handwritten text, but it still struggles with the following:<ul><li>Complex long-vowel mark expressions: The original artwork features powerful, expressive long-vowel marks that extend across multiple lines. This version is unable to reproduce that style of long-vowel expression.</li><li>Mixing handwritten and regular typefaces: It cannot reliably generate sentences in which handwritten characters are mixed with standard, non-handwritten typefaces.</li></ul>Additionally, many of the training images include parts of characters visible alongside the speech bubbles. Since these figures were not mentioned during captioning, the LoRA often generates a woman's head or shoulders when no character is specified. Adding a prompt such as "No figures are depicted in the image" may somewhat reduce this behavior, but completely eliminating it remains difficult.The known limitations of Krea2 itself also remain unresolved. For example, certain commonly used kanji, such as 恥 and 噂, are still difficult for Krea2 to generate correctly.

Concept > Handwritten Speech Bubbles

Initial. (Experimental)This LoRA was trained on MOSS Image Base using exactly the same training images, captions, and training configuration as the Krea 2 version.The results... might actually be pretty good.LORA, MOSS Image Base, dim16/alpha16, adamw8bit, LR 1e-4, cosine 2000steps, 60images.

Files with SHA256: 0ea4af88bce695cfc9afd8556fc0ab6c32ceb4dd061e6361ba4873c9ccaedb27