Everything-Mix v2 [x1024 or HighRes]

Must-use-prompt for anime girl: (1girl, detailed eyes:1.5). Don't always use (perfect hands). Recommended v2 overall with 25 steps and normal res or +960x resolution with 20+ steps.

Important: Here's my tutorial on how to get the best quality possible: https://youtu.be/lF5csBbVIX4 Face fix is a MUST for medium/full body if you use normal resolution. HiRes also an option.

v2_full has hands and quality improvements towards 2.5D+ art, but it is a little worse overall. v2.1c has face improvements and it is very good on closeups (keep in mind sometimes the embedded LoRA can fail, unlike v2). v2.1 has better body, clothes and posture, maintaining the flexibility of making 2D and 2.5D images.

v2.1 seems to work great with HighRes fix.

Samplers: Euler a & DDMI are fast, cheap and good, with +27 steps. Other options: DPM++ SDE Karras, DPM++ 2M SDE Karras. Maybe DPM++ 2M or 2S a Karras (worse) with 16 or more steps (rarely with 11 on closeups), but 20+ is the recommended for normal resolution. Favorite custom resolution for PixAI: 1024x560, 1024x640, 960x610, 960x530. Others: 920x530, 960x640. If you use full body try to re-enhance the model, HiRes fix or use more resolution.

Semi-realistic anime model made merging others. Finally v2 is being released and hands are a big priority for this model. If eyes are not good enough, try to increase importance of detailing eyes. Quality and hand are a lot better than in any other version from v1, and you should now get better results with 20 steps. 30 is still the recommended, you will get higher quality and sometimes better hands, as well as higher resolution. This model doesn't create perfect hands all the time but there will be cases in which you can get pretty good ones. Maybe you need LoRA to polish hands or any other way to change it.

Sampler from both versions? I didn't test it as much, but stay with Euler a, as it gave me the best results. If you test other you may get worse results. Recommended CFGs: 5, 6 or 6.5 for anime, 6.5-7 seems the best for hand generation, 9 and 10 seems to detail more the scenery more, and the eyes in closeups. CFG 5 sometimes creates a different image. Safer generation is 7 but not always the best.

You can experiment with other negative prompts, it is up to you.

If you have worse results with full body, try enhancing the image or increasing the resolution. Also I recommend you to use the option "as a base image", I think the model can have the capabilities to fix itself. If it doesn't work, fix try face fix (even if it's high cost). If all fails, use Stable Diffussion or just try another good model that makes good full body. Try medium view, closeup or any other camera angle if the model generation is forced into full body.

• All versions - v2.1c is v2.1_EX but mixed up with a LoRA to force closeups and improve face/eyes, so now it should cost less credits. It forces the model to do a closeup most of the time, so eyes will be better, but I'd still recommend a LoRA for face and eyes. It still issues with full body, so you have to enhance it. v2_full gives better results with quality tags like (masterpiece) and realistic tags in prompt. "Full" version was done to make hands even a greater priority from all the versions of the models I tried, even if the model gives worse results. So you maybe can use it to fix images. v2.1 is just v2 with a little better face, better body, posture and clothes. I recommend you using my Re;Mix v2 if you want better hands or use it with DarkSky Diffusion LoRA to get better eyes and style. (If I try to change too much the model, it loses flexibility... So I'll keep it like it is)

Some issues I saw: Full body is wrong, but that'll happen with most of the models. Try to enhance the model again. You can learn how with my tutorial: https://youtu.be/BFm2kwJHF3A Hands are much better and you can get very good ones sometimes. Prompts like "perfect hands" makes the AI model go crazy and I think by default works better. Eyes are sometimes unclear depending on the negatives, which can be fixed with LoRAs or forcing the model to detail eyes with prompts. Issues with water, raindrops and tears: Since this model, mixes 2D and 3D sometimes the 3D is problematic. Recommended to fix it with v1.2 or v1.3. I wouldn't use less than 24 steps (Euler a) with 1024x quality, only for closeups without hands.

Unlike v1 that was a 2D model by default, this is a merged model that tries to enhance anime AI art. A mix between 2D and 2.5D... or that was my idea. The good thing of this model is that very rarely I got a bad anatomy issue.

Try to give it different styles and see how it respond. This model tries to be faster, flexible, with good quality and better hands/eyes. I accomplished everything from all my v1 versions. Each quality prompt highly affects this model, so it is sensitive and not limited unlike other models. Keep in mind that.

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

Mixed from my v1.3_style 2D model. I noticed my model gives amazing results in Stable Diffusion with higher resolution and steps, so maybe I should share it... Maybe in the future.

Issues: V2 is a little slow, with lower steps it can show weird anatomy or artifacts even in x1024 res. It works like that because of how I merged the models.