First attempt with the online training!
Geared towards (mainly female-focused) portraits in a subtly surreal illustration style with various types of decorative elements and lots of flowing movement.
NOTE! there are some definite issue with hands as well as fading irises. In any future versions of the lora I will be heavily focusing on fixing those particular two problems. But in the meantime, I recommend adjusting prompts/lora/after detailers accordingly (Sorry for the inconvenience! 🙏)
Some previous epochs can be found over here:
An upload bug unfortunately kept them from being published together for now. E5 on that one sometimes gets slightly better or more whimsical results! If this version is not working for you, please consider giving it a try.
More details on noticed issues! (So far):
I trained this mainly using dall-e generated images, so some of the issues typical to dall-e portraits have carried over. I.e. faded eyes, unfocused eyes, occasional issues with pointed-claw like hands, weird fingers, and long necks. I have noticed tweaking the negative prompt or after detailer is often enough to take care of this. If this does not help with the eyes, I recommend adding a description in the prompt describing eye color and/or which direction of the canvas the character is looking (“looking at the viewer” or “looking left”)
Tips and Recommendations:
Most of the schedulers get good or at least interesting results (my favorites so far are Euler, Euler a, DPM2 Karras, DPM++ 2S a Karras, and DPM ++ 2M Karras) so far the only one that seems a complete no go for most models is DPM++ 2M SDE Heun
Consider adding “long neck” or “elongated neck” to your negative prompt
I included horror, fantasy, fairytale, and sci-fi themed images in the training, so feel free to try out some of these tags for particular decorative elements that fit different moods and genres (e.g. decora_flowers, decora_animals, decora_bones, decora_tech, decora_cyber)
I focused a lot on composition, so consider playing with that in your prompt! (e.g. s-curve composition, diagonal composition, triangular composition, radial composition, foreground silhouette layer, foreground framing, character framing, etc.)
If you don’t like your character’s hair a bit unrealistically floaty, you’ll probably want to add “floating hair, flowing hair” to your negative prompt. The lora also really tool to wavy hair and hair ornaments.
The training really liked the image of someone walking through water, if that shows up but does not your intended image, consider adding some of these to your negative (waves, ripples, decora_waves, decora_water)
Tl;dr—this is an in-progress lora that definitely has some kinks to work out, but I think you will find some of the results are well worth it!
Have fun, let me know if you have any feedback, and I’d love to see any creations you make, even if it’s just to show me flaws in the lora!
-- -- -- -- -- -- -- -- -- --