Sweet-mix
End of the road - After 22 versions and over 2 years, Sweet mix will no longer be updated as of Illust v1.4. Honest thanks to those who stuck around.
Sweet-mix is a text-to-image Stable Diffusion model that is mostly focused on the anime art style. The model itself is a merge of multiple other models, as i don't have neither the time, nor the hardware to train a model. Sweet-mix can be used for both SFW and NSFW images. I think my model is pretty nice, so feel free to give it a try.
MUST-HAVES FOR GREAT IMAGE GENERATION LIKE IN THE IMAGES:
These quality tags: masterpiece, best quality, high quality, detailed, absurdres
This negative prompt: ugly, bad feet, bad hands, bad art, ugly artstyle, bad anatomy, bad fingers, censor, censored, ugly, deformed, noisy, blurry, low contrast, photo, realistic ,watermark, username, text
VAE: SD XL VAE
Good artist styles to use: iwbitu, kesoshirou, shexyo, oroborus, mamiyama, kihou no gotoku dmc, seisui bar, kawase seiki, yuuji (and), shiokonbu, hitmanatee, ashraely, starmilk
And a great imagination :)
Tensor.art link: Sweet-mix - Illustrious/XL - v1.4 | Tensor.Art (free online generation)
Description
Noticed this new Illustrious trend and I decided to try my hand at it. Mixing Pony models with Illustrious XL is a no-go, so I decided to momentarily give up on the Pony line-up and create an entirely new model.
FAQ
Comments (23)
what is new in this update?
Not much really, Its mostly just a new base for upcoming updates. From what i can see, it has better style and hands than v1.3 but has a bit worse feet.
Awesome model! Works really good with controlnet! ggs
Im having problems. My pictures are coming out really purple. Have tried googling the problem but no help. Any help here. I have the included vae and have it enabled
I don't really understand how you're having this problem nor what causes it but I'll try to give you some general tips:
1.Make sure you use a Stable diffusion XL VAE
2.Make sure you haven't toyed with your webui's settings too much, try setting them to default.
3.Try reinstalling your webui.
4.Try other models, maybe its a specific model issue somehow.
5.If you havent used Stable diffusion before, check if your pc is powerful enough to run it. You should have a pc with a graphics card equivalent or more powerful than an RTX 3060.
This could be related to how high your CFG is. I recommend using PNG Info one one of the gallery images and seeing how that looks when you run it.
Is there a reason why you would put pony score tags into the negative for the illustrious version when it doesn't look like Pony has been merged in at all? I severely minimize what I put into negatives because I feel it vastly limits the output in unintended, bad ways. I see everyone else using it but I'm curious if this is more a follow the leader situation and no one really knows why they're adding it -- just because they're told to?
@sansmia Oh, thanks for telling me. The reason why i use pony negatives is just because i forgot to work on a different negative prompt. Its what i used on my previous pony version, carried over for the sake of it. I'll change the negative prompt to be suited for illustrious when i finish working on the next version.
I am not a native English speaker, and the following content is translated by ChatGPT:
I really like this model! It performs exceptionally well in both NSFW and SFW content. However, the only downside is that in my daily use, it often generates images with four toes or extra toes. Even when I use the recommended negative prompts, the issue still persists.
After many attempts, it seems that using a beta scheduler, a non-converging model(like eular_a\sde\etc), and a high number of steps, combined with starting denoising from the third step, can reduce the issue of missing toes.
@SANLVZHETABG Thanks for the support. Yeah, i know about this problem and its been a focus of mine to reduce it ever since v1.1. The model that is performing the best in this case is the latest pony model, a.k.a v1.3.
@Manseo yes,I am currently using v1.3.Thank you for your exceptional work!❤
After using multiple illustrious checkpoints I can say that this model is the one I find myself using the most.
Looking forward to seeing this be more refined and updated i the future
@SnappleRYRY89 I'm glad to hear that you find my model the best. I can say that the next version is around the corner, with it probably coming in around a week from now.
@Manseo appreciate the updates as well
"I've noticed that some parts of the generated images tend to be overly bright. Is there any solution to this issue?"
I really like the characters generated by this model, but I haven't been able to solve the issue of the image being too bright. I hope someone can share a better solution with me. Thank you all!
@linlo Trying other VAEs can help improve the issue of overexposed highlights,maybe
@linlo You can also use iPadaptor to load paintings by real-life artists for style transfer, which significantly improves the highlights on the skin.：D
@SANLVZHETABG Thank you, I will try various VAEs.
@linlo try low cfg below 2,it will reduce saturation,but can get soft highlight!i often use 1.4~1.7
@SANLVZHETABG Yes, changing the VAE can greatly improve the issue of overly bright images. Thank you!，I will continue to try the method you mentioned.
@linlo You're welcome,I've been using this model for a while, and it works great. If you have any questions, feel free to mention me.