AGGA's Utopia —「illustrious-SDXL」

💫 Wanna Try Before You Download?

💻 Try it Online

• 🌀 Launch UtopiaXL online

Just click and generate directly from the browser!

Hi! Here's a quick introduction to the model:

This is a model I created inspired by some of my testers (like CuteMaster CKPT, various LoRAs, and a mix of styles and concepts). NOW SEVERAL 100% NSFW CONCEPTS ADDED!
I recently refined it with the help of Illustrious, and although it was originally meant for personal use — as I use it to generate new styles and concepts — it was also built with several community resources and ended up performing better than expected. That’s why I decided to polish it and share it.

In short: I experimented with many things until I finally got the result I was looking for. It might not be perfect, but I really liked how it turned out.

Hope you enjoy it!

📦 UtopiaXL V2.0

This version includes a bigger number of concepts and tries to rescue some styles that were lost in the previous version (I only added 200 styles since I'm not sure how many illustrious base includes), the ones I've tested work well so there aren't many problems there, it should also include the concepts I added below, but I'm still testing the model to see how many were merged, if it's as I expect, there should be more than 200 concepts (although I only added them to make my life easier, in short... I hope you like it~)

PS: Many of the new concepts added are things already included but improved, such as the hairjob (I actually don't remember almost anything about what I added, but I do have it written down. Perhaps you'll see it in the examples I uploaded or should upload. Most of them will be succubus, so if you see them, those will be the concepts).

⚙️ Recommended Settings:

• Steps: 15–30

• CFG Scale: 2.5–7 (4.4 or 5.5 recomended).

• Sampler: Euler a, DPM++ 2M, etc.

• Resolution: Larger than 1024x1024 to take advantage of its original resolution.

• Recommendation:

  • Clip Skip 1 | CFG 4.5–6 | Greater accuracy in anatomy and clothing, less clutter | Fine-tuning, details, bug fixes.

  • Clip Skip 1 | CFG 5.5–7 | Balanced between detail and stylized expression | General use, complex prompts.

  • Clip Skip 2 | CFG 4–6 | Smoothed, sometimes dull, less expression | Stability on heavy models.

  • Clip Skip 2 | CFG 3.5–7 | Strong style, more saturation or deformations | Artistic effects, exaggerated style.

  PS: Try Clip Skip 1 + low CFG (4–6) if you're looking for precision.

  PS: Use Clip Skip 2 + higher CFG (4–7) if you're looking for more style or artistic effects.

• Recommended Upscaling:

  • Scale: 1.5

  • Steps: 20

  • Upscaler: R-ESRGAN 4x+ Anime6B

  • Denoising strength: 0.35 – 0.5

📌 Note: The VAE is already integrated; no need to load an external one.

🛡️ Safety Labels

This model may generate content of various types. Users are expected to correctly categorize its use into the following categories:

General / Sensitive / NSFW / Explicit

Please consciously add "NSFW" to negative labels if you wish to avoid explicit content.

🏷️ Prompt Tips

Positive Prompt:

sexy_perspective, (masterpiece), (best quality), (ultra-detailed), amazing quality, very aesthetic, illustration, perfect composition, intricate details, absurdres, incredibly absurdres, shiny skin, realistic, 

Others More: anime, cartoon, shiny skin, shiny, slutty, sexy, slut:2.1, dynamic angle, dynamic pose, dark_art_style, abstract element, abstract background, foreshortening, extreme perspective, high resolution, ultra-detailed, extremely detailed eyes, detailed face, newest, best illumination, depth of field, complex background, detailed background

Negative Prompt:

worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, artist name, sketch,  , watermark, logo, patreon username, bad details

📦 UtopiaXL V1.5

This version reduces the strength of the base style and adds some details, it doesn't change much, it makes it easier to see some perspectives.

Things to keep in mind with this version:

• Don't use this model with cutelatte, as this model includes 200% of that data, and that would break the entire model by forcing overfitting, even though it was reduced in version 1.5 (it makes it look unfinished like a scribble). Simply use the descender to force its appearance.

• My idea with this is to include several of my loras so the next version will have a larger flow of data in terms of poses or concepts. I'm still not sure whether to use it as a base or clean up some of the folders, or change the base model by going back to Illustrious 0.1

⚙️ Recommended Settings:

• Steps: 15–30

• CFG Scale: 2.5–7 (4.5 or 5.5 recomended).

• Sampler: Euler a, DPM++ 2M, etc.

• Resolution: Larger than 1024x1024 to take advantage of its original resolution.

• Recommended Upscaling:

  • Scale: 1.5

  • Steps: 20

  • Upscaler: R-ESRGAN 4x+ Anime6B

  • Denoising strength: 0.35 – 0.5

📌 Note: The VAE is already integrated; no need to load an external one.

🛡️ Safety Labels

This model may generate content of various types. Users are expected to correctly categorize its use into the following categories:

General / Sensitive / NSFW / Explicit

Please consciously add "NSFW" to negative labels if you wish to avoid explicit content.

🏷️ Prompt Tips

Positive Prompt:

anime, cartoon, shiny skin, shiny, masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, extremely detailed eyes, detailed face, absurdres, newest, best illumination, depth of field, complex background, detailed background

Others More: slutty, sexy, slut:2.1, dynamic angle, dynamic pose, dark_art_style, abstract element, abstract background, foreshortening, extreme perspective,

Negative Prompt:

bad quality,worst quality,worst detail,sketch,censor, big nose, 
bad anatomy, ugly, worst aesthetic, old, mismatched pupils, dirty face, extra limbs, bad hands, bad fingers

📦 UtopiaXL V1.0

My baking method is a bit more involved than just picking models and merging them with Supermerge (although I do use Supermerge too).

I usually convert models into folder format and manually replace the UNet, while sorting or merging the text encoders—assisted by scripts, of course; I’m not a masochist.

Naturally, the easiest way to control the base style is by creating an external LoRA, baking it in heavily, and using it as a control layer… though I went a little overboard with this version and ended up overpowering some other elements. I’ll be fixing that in the next few days.

⚙️ Recommended Settings:

• Steps: 15–30

• CFG Scale: 2.5–7 (2.5, 4.5 or 5.5 recomended).

• Sampler: Euler a, DPM++ 2M, etc.

• Resolution: Larger than 1024x1024 to take advantage of its original resolution.

• Recommended Upscaling:

  • Scale: 1.5

  • Steps: 20

  • Upscaler: R-ESRGAN 4x+ Anime6B

  • Denoising strength: 0.35 – 0.5

📌 Note: The VAE is already integrated; no need to load an external one.

🛡️ Safety Labels

This model may generate content of various types. Users are expected to correctly categorize its use into the following categories:

General / Sensitive / NSFW / Explicit

Please consciously add "NSFW" to negative labels if you wish to avoid explicit content.

🏷️ Prompt Tips

Positive Prompt:

anime, cartoon, shiny skin, shiny, masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, extremely detailed eyes, detailed face, absurdres, newest, best illumination, depth of field, complex background, detailed background

Others More: slutty, sexy, slut:2.1, dynamic angle, dynamic pose

Negative Prompt:

bad quality,worst quality,worst detail,sketch,censor, big nose, 
bad anatomy, ugly, worst aesthetic, old, mismatched pupils, dirty face, extra limbs, bad hands, bad fingers