Hassaku aims to be an anime model with a bright, distinct, slightly 2.5D anime style.
Join my Discord for everything related to anime models and AI art.
If you’d like to support my work, you can do so on SubscribeStar
Every bit helps me continue making AI models and even an AI VTuber, which is currently in the works.
How to Use
You need to use ComfyUI or Forge Neo.
20-50 steps, CFG 3-6, (i like to use 25 stepts with cfg of 3.5 and er-sde as sampler).
- Put @ in front of artist tags
- Use lowercase for tags and spaces instead of underscores.
- Recommended positive: "masterpiece, best quality, score_7", but since masterpiece is altered, in some cases masterpiece is enough
- Recommended negative: worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts"
- When using a tag that is different between Danbooru and Gelbooru, prefer the Gelbooru version.
Supporters
Special thanks to my supporters:
pttcot, SETI, Kodokuna and dataset captioners mariotheknight and tamashiicolle
Recommended Resolutions for Anima
Here are some resolution suggestions for Anima:
1536 × 640
1344 × 768
1216 × 832
1152 × 896
1024 × 1024
896 × 1152
832 × 1216 — most recommended
768 × 1344
640 × 1536
Larger resolutions are possible.
Note: It is a Diffusion Transformer (DiT) model, it is less restrictive on sizes in compare to a Unet like Illustrious, you can use bigger resolutions.
Version and License Info
V0.1:
A small-scale style training test, intended to see how adaptive Anima is.
V1:
Test for character training with nte characters and femme soule (pukey goddess shot trick). Better use of the style training set.
V1 Style:
Focus on a consistend style. Revert training of nte characters and femme soule (pukey goddess shot trick), because of forgetting issues of anima itself.
License: Anima
Description
Anima has a strong issue with forgetting concepts during training and I was not able to overcome that problem so far. For example, V1 forgot Naruto’s orange outfit. Because of that, I reverted the training on NTE and Pukey back and focused my efforts on making the style more consistent.
The model first received pure style training, similar to version 0.1, but more refined.
On top of that, a refined style LoRA was merged into the model.
Both the style training and the LoRA were trained using natural language and tags, so the style also transfers when using natural language prompts.
This is a strong style-focused model, so artist tags will have a weaker effect as well tags like "realistic" or "no outline".