agartha_xl

Greyhawks/agartha_xl

agartha_xl is a text-to-image model from Greyhawks. Use it when you need an image from a text prompt. It is set up for diffusers. The card lists the license as cc-by-4.0.

Downloads · 30 days

12% of all-time downloads

All-time downloads

341

Public

Repo size

230 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors212 MB · 92%

At a glance

Task: Text-to-Image
Library: diffusers
License: cc-by-4.0
Access: Public
Created: Jun 22, 2025
Updated: Jul 14, 2025
SHA: d7256943

Try a prompt

Base models

stabilityai/sdxl-turbo

Datasets

semiosphere/innersun_dataset

Model card

From the Hugging Face model README

Agartha

Model description

Have you ever seen Agartha?

I bet you have heard and read a lot of it, but images?

Nope! It's always the same fake fantasy RPG, but then how come the inner sun? How come the hollow earth?

The most important part is not having the biased "sky on top" as usual. like, if agartha is in the center of earth, there can not be a blue sky on the top part of the image. then the inner sun, the sun cannot be on top neither near bottom (as a sunset), it's the clouds which will give the effect. there is an image of a night too, but it's the inner sun fully covered in clouds.

Recommended for 1:1 dimension. Read Edit section below.

We also tried other toroidal cities, such as Sigil (from Planescape/D&D,) while it is not 100% accurate, the result was quite interesting.

Positive prompt:
(highres,high quality, circular nebula,,circular blur), sigil,
D&D, the vertical singlering torus toroidal city,
fantasy, magic aura, the city of portals,
Planescape setting, against a flat star lit space,
mediaval magical houses and trees and taverns and towers all
around the ring and planted in the top with ceilings downward
to center, arcane star in the center,
(smooth detailed 3d, professional rendering,)

Negative prompt:
(lowres:1.2), low quality, bad quality, ugly, 


Base Model: juggernautXL_ragnarok
Lora: Agharta - 0.7
DPM++ 2M SDE Karras, cfg 5, steps 30

(circular motion blur), agartha, the high-fantasy world of agartha,
 the inner world where houses and forest and flora and fauna are
all around in the circles,
(and the inner fullmoon in the center),
 (ultra high definition, high octane, professional 3d rendering)

DPM++ 2M SDE Karras, CFG 5, Steps 35.

We switched the base model, insead of SDXL 1.0, we picked "Juggernaut - Ragnarok", and things improved a lot. Here we added "circular clouds", and finally the model was able to set a few houses in the sky, as it was supposed to be. No more extra sun "in the sky" either. Model could understand that there's only a single sun, which is the inner one. (it was trained with that sun being called inner sun all the time.) And those "planets" or "balls" are props like houses and/or anything else, which the models does not know what to do with, and hallucinates those nice flying things. Which is not so bad as we're dealing with fantasy realms, right? We even got a nice Agartha logo. (There is not a single text or watermark in the training images.) I gave it a go with 40 steps, and while more houses were correctly placed in the sky, more "holes" appeared as errors. So we'll stick with 30 - 35 steps.

(highres,high quality, circular clouds,,circular blur),
agartha, the high-fantasy world of agartha, the inner world where
 houses and forest and flora and fauna are all around in the
circles, (and the inner sun in the center),
(ultra high definition, high octane,
professional 3d rendering)

DPM++ 2M SDE Karras, CFG 5, Steps 30

Still on "Juggernaut - Ragnarok version", we decided to lower the Agartha LoRA to 0.7 and then 0.6. We lose the houses planted in the sky, in exchange we got some quite nice images, in medieval style. It was a good surprise seeing things gaining wings or going flying without any mention at it.

(highres,high quality, circular clouds,,circular blur),
the fantasy inner world of agartha, kings and kingdom,
 horses, knights celebrate on the streets,
castle, towers, farms, river, a single centered
image of the sun,

DPM++ SDE Karras, cfg 5, steps 25

Again on Juggernaut - Ragnarok as base model, we set LoRA weight at 0.7, we tried to lecture the model in vain (we knew it would not help, but who knows?) And while not getting the underworld tunnel, it generated a nice image which still fit more or less in the theme. Quality improved a lot too. There are much more generated like this, and all of them look good. Some even get the desired terrain instead of sky.

(highres,high quality, circular clouds,,circular blur),
agartha, the high-fantasy underground world of agartha,
 the inner world where houses and forest and castles and
 towers and flora and fauna are all around in the circular tunnel,
(and the inner sun in the center) and houses and forest far away
around the sun, insane depth towards the center of the planet,
underground realm, (ultra high definition, high octane,
professional 3d rendering)

Edit: I've notice it generates artifacts and that there are some flying things. The flying things are not bad, but they are actually error by SDXL, which does not understand how to render houses and buildings sitting on the sky, as in upside down. Same about the earth "rings", it glitches and renders nothing, and those things keep flying. I decided to give it a go with samplers which can deal with more complex instructions, such as DPM++ 2M SDE Karras, and ta-da, no more excess of artifacts and things tend to be placed where they should. Yet, it's not perfect, and they need high amount of steps such as +30. It helps adding "circular blur" or "circular motion blur", it won't create a circular blur, but it helps the model to understand how to render things. It won't ever understand, but according to its larger dataset, having a circular motion blur, makes it match better the model's internal statistics, which lessens the amount of artifacts (they are caused because the model does not know how to scale things properly.) There is also the fact that we are dealing with huge depthness, even Flux struggles with such perspective, and in SDXL we might have reached its limits? Then I added two extra newer samples (they are the first two images now) to show how we managed to somehow improve it.

How deep are we talking about? There a sample, where the quality dropped a lot, yet we can see how far it goes, and all those things near the sun, are those which are turned into to planets or flying things. Here we can see that the model cannot handle going that further while keeping things smooth, so tons of artifacts and lowres texture appears. Mind you, it's not a REAL depth, it's mostly pixels, which must be scaled and placed in unusual ways. Anyways, here towers are well placed, there's that hole, which is a blue ring in between terrain, at least near the sun, there is a visible settlement, although there's no real ring around the sun, it's whether more houses or mountains. And in the case of mountains most of them were ignored, and such thing also prevent the full-tunnel effect.

Error: Generating nothing

Here I set the CFG to 7 with DPM++ 2M SDE Karras, with 30 steps. Forcing the CFG attains nothing, as it just make the image model glitch as seen in the image below. Those "flying saucers" are nothing else than houses, which bug out the model internal statitics. And they are the one which are then reshaped into planets. At the same time, textures are cleaner, and mostly vegetation is lowres. In our case here, the model is not using most of its dataset, but only ours here. Then it uses as base whatever it learned, and it was few, and due to their addiction to oversmooth, now that they need to really rotate things, all the lowres texture showup and they do not even know how to oversmooth it again. lol Those errors are mentioned above in the Edit section.

Note: that's why people need TO STOP traning images as amateurish, the subject always centered is a typical ai-image, and the excessive amount of sun in the sky for beauty (not realism, because reality is ugly,) makes it impossible other things such as Agartha. hehe Yes, you smexy girl is NOT realistic. Neither your 6-pack male, all of us are ugly and plump. :P

Soundtrack

"Agartha"

Dataset

Well, so in order to share what we got the best, I'm offering a dataset called "Inner Sun". It won't be all about agartha. But they are not tagged, but properly captioned via civitai.

That way all of you can join us in the journey of breaking the models.

Good luck!

Trigger words

You should use agartha to trigger the image generation.

You should use inner sun to trigger the image generation.

Download model

Weights for this model are available in Safetensors format.

Download them in the Files & versions tab.

Task: Text-to-Image
Library: diffusers
License: cc-by-4.0
Created: Jun 22, 2025
Updated: Jul 14, 2025