CitrineDreamMix

Overview

CitrineDreamMix is a highly versatile model capable of generating many different types of subjects in a variety of styles. It excels at creating beautifully detailed images in a style somewhere in the middle between anime and realism. It also has a strong focus on NSFW images and sexual content with booru tag support.

This model is a merge of two of my favorite models: AbyssOrangeMix3 (AOM3) and NeverEnding Dream (NED). I take no credit or ownership for any of their work.

AbyssOrangeMix3 is great at making intricate anime-style illustrated images, but it also has a tendency to add too many little details and stray a bit too far from realism. It handles fictional characters and NSFW content very well though, even more niche stuff.

NeverEnding Dream is great at making clean and realistic images, but it also has a tendency to simplify and limit details too much, which can make images appear somewhat flat and basic. It can also struggle with some more specific fictional characters and NSFW scenarios, especially those that stray away from realism.

The goal of CitrineDreamMix is to find a happy middle ground between AOM3 and NED to benefit from the best aspects of both. It has a more realistic style than AOM3 and a more anime-inspired style than NED. That allows it to retain the anime-style physical features like large eyes while also retaining accurate and smooth details more in the realm of realism.

Below are a couple images that compare CitrineDreamMix with AOM3 and NED to showcase the difference in styles. I have also included some other larger size comparisons in the example images. I tried to keep the prompts fairly simple to show their normal style since all of the models can be steered more toward realism or anime-style using prompt keywords. They all produce beautiful images, so it's just a matter of the subject matter and personal preference.

Model Comparisons:

Prompts

For prompts, my main recommendation is to make use of booru tags since both AOM3 and NED are trained to use those. If you are using A1111, I highly recommend the Booru tag autocompletion extension since it makes figuring out the exact format of the tags much easier.

If you want a more realistic style, use "realistic" in your prompt, and if you want more of an anime style, use "realistic" in your negative prompt. You may need to increase the weight to see a noticeable difference.

For negative prompts, I use some of the popular negative embeddings such as bad_prompt_version2, EasyNegative, bad-artist, and bad-hands-5. However, I have not done extensive testing to see how much of an effect these have, so don't think they are necessary.

Settings

Disclaimer:

I'm not a Stable Diffusion or AI image expert in any regard, so take all of my suggestions with a huge grain of salt. These are just some of the best settings that worked for me after some limited testing. I would recommend you play around with the settings and find what works best for you.

Sampling Method:

I found DPM++ SDE Karras to work the best followed closely by DPM++ 2M Karras, Euler a, and Heun. However, almost any method can work just as well depending on the settings and desired result.

Sampling Steps:

I found around 20 steps to be good for me, but anywhere from around 10 to 30 steps seemed to produce similar results. Your results may vary depending on the sampling method used.

CFG Scale:

I usually just stuck with a 7 CFG scale, but anywhere from 5 to 9 seemed to produce almost identical results. Around 10 and over sometimes resulted in some weird artifacts and messed up proportions, so I would be careful setting it too high.

CLIP Skip:

I found using CLIP skip 2 produced slightly better results than CLIP skip 1, but both work well. CLIP skip 1 seemed to produce a more realistic style, but I didn't do a ton of testing, so that could have just been a coincidence.

VAE:

I liked Waifu Diffusion's kl-f8-anime2 VAE the best, so that's what I included in the baked VAE versions, but the standard vae-ft-mse-840000-ema-pruned VAE is also very good. The orangemix VAE seemed to produce duller colors for me, so I would only use that if you are looking for more realistic toned down colors.

Upscaling:

I upscaled all my images by 2x using the Hires. fix setting during txt2img.

I like the 4x-UltraSharp upscaler (can be found here along with other upscalers), but there are many other good options like R-ESRGAN-4x+ and even the Latent ones. This is very much a matter of personal preference.

I set the denoising strength to around 0.5 to 0.6, but you could go more or less depending on the upscaler being used and how much you're upscaling by. Also, if you are using the SD upscale script in img2img, you probably want a lower denoising strength around 0.2 to 0.3.

Changelog

June 16, 2023

• Added v2 models to use latest versions of AOM3 and NED. See version notes for improvements and differences.

March 11, 2023:

• Added v1.1 models to fix issue with merge size.