Important: Use VAE for the versions that need it or you will get a bad generation.
I included a new baked VAE in the model, use it! Recommended 30 steps and CFG 11. Try different samplers if you want but my favorite for this model is DPM++ 2M SDE Karras and sometimes DDMI, Euler a and DPM++ 2M Karras. You can try to lower steps a little for less quality (ex: 28), but CFG is important. Lowering it can mess up the generated image.
Model made by teasan. It is one of the best semi-realistic anime models with the best style/backgrounds I've seen, but it is limited. I'm uploading v1.21 as it is the last version. If you use versions that doesn't have VAE it will look less vibrant and with little contrast. Test freely with different samplers, but the recommended by the creator is DPM++ 2M Karras.
As negative prompts, I leave by the default ones by the creator, it gives good results. Be careful if you test others. V1.21 seems to have improvements in facial generation compared to v1.2. It looks more "experimental" than v1.1.
Recommended prompt tags by creator: absurdres, highres,
Recommended by me: (best quality), (masterpiece), (ultra-detailed:1.2),(photorealistic:1.1), ultra high res, extremely detailed
I think it is one of those models that doesn't need too much quality tags, so use a few.
Recommendation from me: Use the model outside PixAI with VAE and custom embeddings. Also I recommend to refine the result with another model.This model doesn't seem to create perfects hands most of the time, so you may need a LoRA for that... and maybe you can play with embeddings. You can also try something like: (low_contrast), (low_saturation), in the negative prompt. I think it helps a little in PixAI.
I also recommend using LoRAs with 0.5 to 0.7 height so you can polish details if you use lower steps. Stable diffusion users: Recommended clip skip 2.
Recommended resolution for PixAI: 1024 or 768 with high res fix. Also don't forget to use highres fix for your images to x1. If you do it right, it is free!
