Sign In

[XL] Swirling style

5

64

5

2

Updated: Jun 1, 2025

stylepsychedelicsurreal

Verified:

SafeTensor

Type

LoRA

Stats

64

5

1

Reviews

Published

Jun 1, 2025

Base Model

SDXL 1.0

Training

Steps: 11,000
Epochs: 25

Usage Tips

Clip Skip: 1
Strength: 0.8

Trigger Words

surreal
psychedelic
underwater

Training Images

Download

Hash

AutoV2
6EB30E3A46

A cool style snatched from Dall-E 3/Bing while looking for decor concepts.

It will engulf your scene into a lot of ultra-crisp undulating lines, beads and glowing flows of matter. Best with 2.5D/semi-real/photorealistic models as this is the original style.
Dataset mostly contains 1:1 images and some landscape ones as this is a limitation of Bing creator, so it may have some hiccups with vertical images.

I trained it on Illustrious and Flux as well but i didn't get this "wahoo" feeling there - i'm not sure how to train those proper and having the results shown here is some matter of luck to me.
So for now i'm only publishing the XL version.
Original dataset + bing prompt provided, feel free to try and improve :)

-- Important: you are free to reuse this model, but if you use it in a merge or retrain it, please provide a link to this model's page or mention me. Thank you ^^

Usage guide and tips:
Recommended weights:
- Soft: 0.5
- Normal: 0.75
- Heavy (with risks of distortion): 0.9-1

Best for solo portraits, duos depending on image dimensions

Bodes very well with fire, feathers and marine life. I tried lightning too, it's good but less impressive.

No trigger word. Just use surreal, psychedelic, underwater + lora in prompt. I got pretty nice results with very simple prompts.

ADetailer may be needed, especially with high weights.

Known caveats and limitations:
It has more side view than faces, and poses are a bit stereotypical, you may want to use ControlNet for custom stuff.

Style tends to be semi-real/photorealistic, untested but i expect it not working that well with anime models.

Fingers may get "welded" with the background sometimes.

You -do- want to use ControlNet for any face accessory, like masks.

Camera angles like from above, from below may not work "as-is".