Hassaku aims to be a anime model with a bright and distinct anime style.
My Discord for everything related to anime models and art.
If you'd like to support my work, you can do so at: SubscribeStar, every bit helps and is truly appreciated!
____________________________________________________________
Supporters:
Thanks to my supporters tamashiicolle, pttcot, SETI and Kodokuna
____________________________________________________________
Using the model:
The model was trained primarily on images with minimal disruptive elements, such as floating text, logos, speech bubbles, or signatures. If any of these elements appear in an image, please include "signature" in your negative prompt.
Danbooru Metadata and franchise tags are not used in training, so please avoid using them in prompts. Metadata such as "highres" and franchise tags like "re:zero kara hajimeru isekai seikatsu" are not used in training.
A CFG scale between 3 and 7 is recommended. For best results, use CFG 6 with the Euler a sampler.
Refer to the example images, to understand, how the model should be used.
The model is designed to be simple and straightforward to prompt. Use the following tag order:
Number of characters
Character Tags
All remaining tags
Example: 1girl, rem \(re:zero\), standing, masterpiece, upper body
Here are some resolution options for SDXL:
1536 x 640
1344 x 768
1216 x 832
1152 x 896
1024 x 1024
896 x 1152
832 x 1216 (most recommended)
768 x 1344
640 x 1536
Lora use:
NoobAIs Loras working for the most part better in compare to Illustrious loras
Versions 2 and 3 is less compatible to most loras trained on Illustrious or Noob, because it got further trained away of both. Style A is still the most compatible for it.
______________________________________________________
Version and License info:
Below V1 merges ANIMAGINE XL 3.0
Version V1 use Illustrious-XL & WAI-NSFW-illustrious-SDXL with additional training
V2 is trained on its own and don't include any extra merge. Base was V1, it is trained to include newer or missing characters. It was also used for further training tests
V3 is a merge of Illustrious-XL and V2.2, to fix some issues and was also trained to include newer or missing characters
All Models using the Fair AI Public License 1.0-SD license.
Description
Because of some selective training, it seems that it got a bit unstable, especially hands have more problems. I'm not satisfied with the result, but it is overall slittly better then before.
FAQ
Comments (20)
Notes on V0.3: Seems a bit less stable and it have a error, if you use ":3", model goes into the abyss. Working on that and on more expressions.
It is, like i said already, far from finished.
No merges went in, to stabilize it. No add difference or other technique.
I basicly take half baked wares up here. My goal is to create a new NAI, but with only one 4090, it takes time. I want it first to be fully anime, with a high responce on danbooru tags. It is now slowly transitioned to it, you don't need to put anime in positive or 3d&realistic in negative for the most part anymore.
It is for now a bit to much 2.5d, i want it more in between sudachi and hassaku.
I tried both v2 and v3, both are giving me very high contrast images. Trying to use some of your example images prompts it still results in high contrast images, using same settings. Do you have any suggestions ?
Thank you for continuing to share, even this early on. It is easy to merge your work in ComfyUI with other checkpoints so I use Hassaku like a really big LoRA for now! Looking forward to seeing where this goes.
@digger73 Have you tried using the StabilityAI offset noise LoRA with negative weight? In theory that should reduce contrast, although I haven't tried it yet. You can find it on Huggingface: "sd_xl_offset_example-lora_1.0.safetensors"
@digger73 Do you mean, it have a higher contrast then the examples? or the same contrast? I train it to have a relative high contrast in compare to other anime models
@Ikena It has a higher contrast than the sample pictures. I cannot generate the same results. It seems that if I reduce the CFG to 5.5-6 I get the same kind of look as in your samples, with less contrast.
What is the training data comprised of? Danbooru images, or have you cherry picked results from SD 1.5 and put them into the training data? Just curious
over 140k filtered and cleaned ´"masterpiece" danbooru and gelbooru images with around 1k cherry picked SD1.5 images.
u could also include dalle3 images in that mix as well its only available on bing image creator but some peeps appear to already using their advanced qualities to make LORAs @Ikena
@Stablediffusionlover I can, but one concept have arround 1k to 20k images in my dataset. Making effort to generate hundred of dalle3 images is a bit much
DALLE-3 is meh at anime anyway, it's a jack of all trades model and a really good one at that. I think training it solely on human data is most ideal but just my opinion. I think some SDXL models have been clearly fine tuned on SD 1.5, especially noticeable in the background details and in some models hands but that less so. Also, I think trying to imitate DALLE-3 with SDXL wouldn't bare any fruit. DALLE-3 is simply better at understanding language at a architectural level probably, we would need better more advanced base models with better CLIP model etc. Probably to get anywhere near it I imagine, and even if we did get better base models good luck running them locally for now. But more powerful base models would be welcomed even if most people couldn't run them locally, in the LLM space there's 70B models that even the most top end computers can't run but are competitive with the top models and you can always run them in the "cloud"
Can I run this model on A1111?
why shouldnt you?
@mech4nimal I'm just curious, since this is my first time using XL model
Yes, but you might need to update
@Iffymiffy Ok, thank you
One more question, can I run this on Low VRAM GPU? Like 1650?
@Samapoke Possibly wont be able to on low VRAM, or at least will take a while
@ZeroUnits Thanks
If some wants to know, how my progress is going. I post updates on my channel "what-i-do" here: https://discord.gg/sd-citrus-models-1072156418402156624
Update tomorrow