THIS IS A VERY EXPERIMENTAL MODEL. Things that I see wrong will be fixed in v1. Important: For full body, you have to re-enhance the model, HiRes fix or use more resolution. 2.25D model more oriented to N-SFW and Hen-tai, but it can still provide SFW (normal) generation. Merged with other models. Since I'm experimenting a lot I'm going to name it v0 instead v1. My intention is to keep experimenting until I find a good version for v1. I recommend face fix for full body or you will get worse results. Enhancing may also help.

⚠️ Here's my tutorial on how to get the best quality possible: https://youtu.be/lF5csBbVIX4

Keep in mind that if you really want a better image you need to increase steps A LOT, but in PixAI it is better if you learn how to enhance the image. For closeups you can use lower steps. LoRA recommended for all versions since I didn't try to fix hands yet.

Good samplers I've tried: DPM++ 2M Karras with +16 steps. DPM++ 2M SDE Karras with +16 steps. You can use +11 steps in closeups with those. But probably 20+ steps is the safest for all kind of generations, like full body. 2M sampler can be similar to Euler or DDMI, 2M SDE is more interesting and changes more over time. 2S a Karras might also be worth trying.

v0.2 is more flexible and it looks good on many samplers, but it can burn more easily with lower steps. For that reason and for the moment, renamed to v0.2_EX. Important: Since it is a very new model that is hardly experimental, I didn't test negative prompts enough. Test it yourself if you want.

Current versions: v0_style (renamed from v0_EX): The initial version. Focused on N-SFW and also face. I'm not sure if hands are good. Better than it seems initially. v0: The final version I decided to use, similar to v0_EX but with another way to generate the images. Everything-Mix v2 mixed up with other models. v0.1: Very little change from v0. Theorical improvements in hand and N-SFW generation. v0.2 N-SFW+: Hard changes from v0.1. This version mixes up my lastest version of Everything-Mix (2.1_EX) with other models. It looks good enough so I think I am going to use this version to create v1. Maybe I'll try to mix it up again to see if it is better. Hands are better than expected but not always perfect.

If you want to force less N-SFW try something like (cleavage, N-SFW:1.5) in negatives but v0.2 hardly forces N-SFW.

If you have issues with the model, try to test with different negatives from other models.

As always, if you improve resolution and steps you may get better results. I still didn't test if custom resolutions are better. Favorite custom resolution for PixAI: 1024x560, 1024x640, 960x610, 960x530. Other: 920x530, 960x640.

Best samplers in general for Everything-Hen v0: DDIM, DPM++ SDE, Euler a

Best samplers I tested yet in general for every model: Euler a and DDMI = Good quality and better composition (in my opinion) DPM++ SDE and DPM++ 2M = Cheaper, faster and good quality DPM++ 2M SDE Karras = Very good quality in some prompts but requires more steps, so it is the most expensive for PixAI. DPM++ 2S a Karras (& 2S) = It works in rare ocasions. Worth to try in closeups.

Issues: If closeups are too close, it doesn't make a good image. Always try things like "1girl". To help the model try quality tags, the position of the body and the view in the prompt.

I recommend you to enhance instead using Hires fix that it is a waste of credits. To get a similar result than highres fix in Stable diffusion, you can use the resolution 640x448 and then enhance. This will make your generation worse, but cost will be cheaper.

Off-topic: If you like AI music I have an AI band. Please subscribe! https://www.youtube.com/@ForTheSkies