Everything-Mix 2D v1.3 (And more versions)

Prompts I like: (1girl, detailed eyes:1.5), beautiful eyes. FACE FIX is still recommended. Recommended resolution: 960x or 1024x. CFG: 6-11. I recommend not using high priority for v1 models because it will cost you many credits. You can adjust numbers of negatives. If you want more contrast but worse backgrounds, try (bad-artist-anime, by bad-artist:1.5) in negative prompt.

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

NEW! Everything-Mix 2D v2 now available: https://pixai.art/model/1690068301903299164 NEW! Everything 2D v1.3_style is now 2D v1.1_base.

First model I did merging models. High quality anime style. Recommended: v1.2 and then other non-beta versions. Even if it is v1 the model is STILL ON BETA. Feel free to test it with new negative prompts (let me know if you think there are better ones). It's working better in Stable Diffusion for me, as hands and composition get better with embeddings. Recommended the use of LoRA for hands in PixAI.

New version 1.3: Better hands, better eyes and more 2D. I had to do it this way to fix the face. Even if it lost some identity now it is more similar to Everything-Mix 2D v2.

This model seems like it doesn't need many quality tags. Some quality tags I see working a little: 1girl:1.5, dynamic pose, detailed eyes. Don't add too many like: masterpiece, 4k, ultradetailed, high res... This model doesn't need it as much. 2-3 will make it HD, it is better to increase negative prompts for bad generation. Increase your quality tags by () if you used a least one. Also don't try high semirealistic prompts.

Also, something I noticed from my v1.2 model: Less CFG (6): More anime style. More CFG (11): More semi-realistic anime style. Use this to control semi-realism instead prompts. I recommend 6, 7, 10.5 and 11. CFG 11.5/12 can work sometimes amazingly for hands and with other negative prompts. Safer generation: CFG 7 or 11/10.5. If you want to use v1.2 with HD VAE go here: https://pixai.art/model/1664890692692122022/1664986276374794510

Recommended LoRA for hands, eyes and enhancing/more details. [NEW] Now that custom resolution is here try 960x resolution. I use that because then I can just upscale to x2. More resolution is recommended but not necessary. Steps are more important.

Issues with all v1 versions: Full body. Try to increase resolution or re-enhance if you want to fix it. Sometimes bad eyes or hands, but you can fix eyes and hands with negatives, so negative prompts can be improved.

For the best quality possible I recommend high res fix (Never use this option as default, it costs a lot) or 1024x quality. The more resolution, the better. This is a high resolution model so it will be better with 1024x than 768x resolution. For PixAI, I recommend highres fix to x1 res. because it is free. Configure it on settings with x4 batch.

What you can also do to get an amazing generation for 0 credits: Generate a image with Everything-Mix v0.5, then fix it with Everything v1.2 (as a base image) with 30 steps. Beta versions are there for enhancing.

You can use any sampling method, it looks like it works with everything (20 or more steps, there are a few that can generate with less than that). My favorites for this model are Euler a, DDMI and DPM++ 2M SDE Karras. Have in mind some model versions would require more than 20 steps. Try to not show hands as it will show better results overall. Try something like "perfect hands" if you want decent hands in V1.2+. Sometimes it is best to not use it.

v1.3_hand theorically has better hands compared from v1.2. It has been changed very little.

Best samplers I tested yet in general (best to worst) with 30 steps: Euler a | DPM++ 2M SDE Karras | DDMI | DPM++ SDE | DPM++ 2M | DPM++ 2S a Karras

Comparison between versions: v0 to v1 https://pixai.art/artwork/1664692839678748266 (v1 improved even more) v1 to v1.1 https://pixai.art/artwork/1664870231705030661 (v1.1-Final is now v1.2-EX) v0.5 to v1.2 https://pixai.art/artwork/1665022390419102407 (v1.1 is now v0.5) VAE: https://pixai.art/artwork/1666542801431114911 Negative prompt: https://pixai.art/artwork/1666664414971792475 CFG from V1.1b: https://pixai.art/artwork/1671355728772532902

• Recap - Renamed v1.3_style to v1.1_base I won't upload v1 here because I think v1.1b is the improved version. v1.1b it is a new version released after learning what I've done with v1.2, it has improvements in hand generation compared to v1. V1-V1.1 seems that works better in DDIM and other Karras samplers with 30 steps, and they are a little worse in Euler_a. Beta V1.2-EX and EF were going to be V1.1 but they were too experimental and maybe you can get fine hands with embeddings. I like more EX than EF. Those version can create weird eyes so you need something to fix it. It looks that if you add many prompts, hands get worse. v1.2 sometimes has some improvements in hand generation compared to v0.5/v1/v1.1. All V1.1+ (v1.1 to v1.2) versions tend to use more steps. It can be fine at 20 but it will be better with more. Beta v1.2_hand is a version that tries to improve hands from v1.2. You can expect some changes. v1.3_style(EX) was the version I used to merge a model into v2 and it might be the official v1.3, but hands wasn't always on point. It might try to fix that in the future, but hands are the hardest thing to make good.

V0.5, V1 and V1.1 create decent N-SFW (normally has bad hands). v1.2 experimental versions doesn't create good N-SFW imo, but sometimes the result is fine. v1.2 creates good N-SFW but sometimes it requires more specific tags and steps.

Due to new explicit content rules, add or remove "(N-SFW, cleavage:1.2)" at the end of negative prompt.

Same model with a different VAE, with more saturation and sharpness: https://pixai.art/model/1664890692692122022

Since I don't know much from model merging and I'm still learning, I'll keep the model private for the moment. Please don't ask for downloads. All the versions are fp16, most of them pruned.