I wasn't particularly interested in this character, but I was watching the movie again recently and decided to make a model for her since I haven't seen any yet. Dataset was, of course, limited. LoRA can replicate her school uniform just fine but not her amusement park outfit. It does have some trouble generating her freckles in combination with the vertical lines every character has on their noses in this movie. Her nose sometimes just ends up looking like a blurred mess. Which brings me to my experiments with network dim/alpha since I'm still learning this stuff...I still find training at 64 network dim to create more accurate and more consistent generations; of particular note with this specific LoRA, the 64 dim model I trained was significantly better at generating her freckles and nose lines without the image being too blurred. However, the large filesize (relative to the other LoRA posted on this site) bothers me.I trained an 8 dim model with the same dataset... it was practically unusable due to the freckle/line problem as too few details were retained. The logical conclusion would be to train somewhere in between... but I got tired of poking around with this one especially since I don't feel any particular draw to this character.Instead, I resized the 64 dim version into an 8 dim version using the resize tool... in my opinion, the downsized version performed noticeably better than the version I directly trained on 8 dim.

Sahara Miyoko - Koe no Katachi (A Silent Voice)

saharaV1.safetensors

v1.0 64 dim

v1.0 8 dim resize

Creator: <a href="https://civitai.com/user/HH840034" rel="ugc noopener noreferrer nofollow" target="_blank">HH840034</a>

Files with SHA256: 962c51b59aa5c1b03c2c3cd49b7f729c9a3ad21918e475caf03e399e0b5dbc56