Sign In

Prompting for SDXL: How to get the most out of your models!



With Stable Diffusion 1.5 being dominant for so long, most of our prompting experience come from it and with IDEs such as a1111 UI that lets us save styles or re-utilize the favorite settings we have been carrying too much from SD1.5 into SDXL. With so many amazing tools such as control net, inpaining, comfyUI, we have gained some control over the generation of what we have in mind, however he have set aside our prompting skills which are as poweful as the other tools or assets I mentioned before, even as powerful as so many existing loras.

That being said, I will be using a model I brewed "Hentai Mix - Road to freedom" as an example to illustrate all the capabilities that, if I am allowed to say it myself, good SDXL models have. Please note that Hentai Mix is a model to make hentai, that was the only purpose behind it! if you want to create amazing art, photo realistic images, general anime or any other specific content, there are models that specialize on those things, however, if you want to create hentai or general non realistic nsfw content, I assure you that at current date, "Hentai Mix - Road to freedom" on it's version 5 is one of the best sdxl models for it.

TLDR: SDXL models are able to do amazing things that most of us are not using because we are accostumed to SD1.5 models.

The simpler, the best!

SDXL is quite responsive to prompt, for that reason, every word has a great influence in the results. What this means is that when our prompts are to complex, the words that describe what we actually want have to compete and will have less influence!

But then, how do we fix that? for starters, we need to reevaluate the styles we have saved and keep using on our prompts, checking what effects do they have on results. For this article, I am going to use two styles I have saved to compare results when changes are relevant:

  • Quality:

    • prompt: (masterpiece), best quality, highres, 4k, 8k, amazing quality, amazing shading, soft lighting

    • Negative: low resolution, monochrome, bad quality, deformed, bad anatomy, ugly, extra limbs,extra nipples, text

  • Quality(lite):

    • Prompt: HD

    • Negative: bad quality

Let's now see what difference does it makes in our prompts.

This is a Mecha generated with Quality(lite) style:

prompt: Mecha (style: Quality(lite))

And this is same Mecha (same prompt and seed), generated with Quality style:

prompt: Mecha (style: Quality)

As we can see, even when both images have a Mecha, the style changed drastically and sometimes can even cancel out the prompt we entered, specially when we are trying to do certain styles.

To illustrate this, let's see the following example:

prompt: fantasy art by luis royo (style: Quality(lite))

prompt: fantasy art by luis royo (style: Quality)

As you can see, even when we used the same seed and same prompt, the author style was completely washed out by all the quality wordings that were included in the style. This same principle applies for everything we prompt causing some nice effects to be hidden and not noticeable in the results.

Drawback of using little tokens for quality is that sometimes results have defects, this defects however, can be fixed by adding the custom error happening to the prompt, for example, "bad hands", 'Extra leg", etc. By doing this, the result will be affected just enough for the problem to disappear and the desired effect will remain.

Why use LoRa when you can just prompt?

SDXL models are very powerful when we let them work, I am going to list the capabilities I have found by trial and error along with the help of VeerDeer and Lextramoth who contributed to this list.

Please keep in mind that some (not all) of the effects are only visible when the prompt is light enough!

Concepts and effects:

(Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 744428527, Size: 1024x1400, Model hash: 6dbcf28d67, Model: Hentai Mix - Road to freedom v5.0, VAE hash: 235745af8d, VAE: sdxl_vae.safetensors, Clip skip: 2, Version: v1.6.0)

prompt: 1girl, bioluminiscent (style: Quality(lite))

prompt: 1girl, fluorescent (style: Quality(lite))

prompt: 1girl, black light (style: Quality(lite))

prompt: 1girl, psychedellic (style: Quality(lite))

prompt: 1girl, porcelain (style: Quality(lite))

prompt: 1girl, glass skin with mechanical parts inside (style: Quality(lite))

prompt: 1girl, mecha (style: Quality(lite))

prompt: 1girl, cyberpunk (style: Quality(lite))

prompt: 1girl, steampunk (style: Quality(lite))

As you can see, several styles are already built in, you can just test them yourself or come out with new ones, such as ancient egypt, skeleton, zombie, diamons skin, magma skin, etc. Your vocabulary is the limit!

Artist styles:

There are many Loras out there intended to resemble the style of famous artists, however, there's already a miriad of artist styles that are already incorporated in SDXL models. I am not saying these LoRas are useless, they aren't by any means, finetuning models allows us to use the styles even with a lot of quality modifiers and some artists are not included, specially recent ones or digital artists.

The full list of artists used for training is too big to be put in an article, however I will provide some examples or really cool ones.

Note: This list is not SDXL specific but it's categorized and easier to browse.

prompt: fantasy art by Frank Frazetta (style: Quality)

prompt: fantasy art by Ted Nasmith, poison faerie (style: Quality(Lite))

prompt: art by Beksinski (style: Quality(Lite))

prompt: art by Simon Bisley, HD, summoner, wearing robe with a hood, orc (style: Quality(Lite))

prompt: art by Al Capp (style: Quality(Lite))

prompt: art by milo manara, 8K (style: Quality(Lite))

Characters and personalities

In the same way that artists are added by default, there are a miriad of anime characters, game characters, animated movies characters or real life personalities that were also added, some by default, some by mixing with specialized models trained with them. The list here is hard to come by but I have tested the following with success!

  • Anime characters

  • Video game characters

  • Celebrities

  • Animated movies characters

That being said, I failed to generate Vtubers, it seems that those weren't included!

Some examples

prompt: Mikasa Ackerman from Attack on Titan, 8K (style: Quality(Lite))

prompt: billie eilish (style: Quality(Lite))

prompt: princess peach, 8K (style: Quality(Lite))

prompt: princess Jasmine from aladin, the arabian dancer, disney style (style: Quality)

Several checkpoints in one

Just like you can change the art styles, you can also change the general style, like if you were using a different base checkpoint, just with the prompt

Let's use one example to include them all, this are the generation parameters:

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1899012539, Size: 1024x1200, Model hash: 6dbcf28d67, Model: Hentai Mix - Road to freedom v5.0, VAE hash: 235745af8d, VAE: sdxl_vae.safetensors, Clip skip: 2, Version: v1.6.0

prompt: 1girl (style: Quality(Lite))

prompt: 1girl, anime style (style: Quality(Lite))

prompt: 1girl, comic style (style: Quality(Lite))

prompt: 1girl, 3D render (style: Quality(Lite))

prompt: 1girl, photorealistic (style: Quality(Lite))

prompt: 1girl, fantasy art (style: Quality(Lite))

prompt: 1girl, manga style, ink lineart (style: Quality(Lite))

If you see, only the photorealistic style failed to be achieved due to the nature of the model, instead a semi realistic style came out.

Added value of the base model

After knowing all the capabilities of SDXL models, why then should I pick a specific model? well the anser is complex, but in short, the answer is "Because of it's added value". What does this mean? some models have been trained with specific goals in mind, those goals can give them special powers other models won't have. Additionaly, some models excel in their specific style, like we could see before, "Hentai Mix - Road to freedom" is unable to create photorealistic content, but other models can.

The special power of "Hentai Mix - Road to freedom" is the ability to recreate sexual acts without the need of LoRa, that means you could use all what we just learnt and use it for nsfw content. Since I want to keep this article to artistic nudity at most, I won't be showing examples as those can be found in the model page, instead, I will only list the capabilities here:

  • Blowjob

  • Cunnilingus

  • Licking Penis

  • Handjob

  • Cumshot

  • Doggy Style (pov)

  • Reversecowgirl (pov)

  • Riding Cowgirl (pov)

  • Titty Fucking (pov)

  • Anal (pov)

  • Fellatio (from side)

Final words

I hope that this article helps you generate better art regardless of what model you want to use, the most important thing in my opinion is to be creative, persistent and always keep improoving, loving what you make! Feel free to ask questions at any time, give suggestions and share your creations!