Type | |
Stats | 267 7 |
Reviews | (42) |
Published | Mar 2, 2024 |
Base Model | |
Training | Steps: 15,184 Epochs: 8 |
Usage Tips | Clip Skip: 2 Strength: 0.8 |
Trigger Words | SaginomiyaIsumi IsumiRaisedSleeve |
Hash | AutoV2 97B17DECA1 |
This is my first go at making an anime character model and it's not a high quality one.
Isumi is a side character from Hayate no Gotoku that I like, but there aren't many quality images available featuring her, so I ended up using mostly material from the manga (usually upscaled).
For that reason, this model is best used with quality enhancing positive and/or negative prompts. If you don't want to create manga-like greyscale images, it is helpful to add "monochrome" to the negative prompt and/or "colored image" to the positive prompt.
The activating tag for the character is "SaignomiyaIsumi", just like the model name. Also I used "1girl" and "solo" for all training images.
I did a little experiment this time and added an additional concept for the character, which is her frequently used gesture of raising her (oversized) kimono sleeve towards her face or mouth. It is represented by the tag "IsumiRaisedSleeve". It's not super consistent due to the lack of image material, and due to a high number of repititions used will sometimes bleed into images that don't have this tag; one way to alleviate this is to add the tag as a negative if you don't want it.
This character almost always appears dressed in kimono or yukata in both the anime and manga. I did tag clothing like "kimono", "yukata", "obi" for the model, so it's possible to make her wear something else, although you may have to be a bit insistent with the tagging for that.
There is a tag "floral print" that was used pretty often for kimonos that have a flower design.
I did my best to tag all training images for stuff like the visiblity of the character (type of camera shot), the point of view of the camera, the direction the character is facing and looking, and the basic character pose (mainly standing, sitting, kneeling in the training images), so as a result, this model is decently flexible with these.
This model was trained on Anything 4.5. I have tested it with various 2D anime checkpoints and it seemed to work decently with most of them (also see showcase images). It can work with 2.5D, but not really with 3D.
I don't have an optimal recommended weight, but I have been using the model with weights from 0.7 to 1 for the testing.