I pick up some environment art from Pangya fantasy golf and use AI to change them into a different look.
Original
Retro Futurism
Saturated Space
Wasteland
Wind Hill
Original
Retro Futurism
Saturated Space
Wasteland
Deep Inferno
Original
Retro Futurism
Saturated Space
Wasteland
All of these were made in playgroundai.com
The use of AI might soon become useful for creating new art to be used by the private servers, e.g. in the launcher or main menu banners. Thereās of course still a lot of development to be done until it will be possible to generate faithful reproductions of all the characters or courses, but what we have now is already promising. Iād love to see an AI art generator trained on Seedarkās art style, for example.
I have done some attempts at rendering some of Pangyaās characters using Nemusonaās Waifu generator, which has been suggested to me by an user on Debug Pangyaās Discord channel: https://waifus.nemusona.com/ Among all the characters I have tried to render, I have managed to get decent results only with Kooh, for all the others I believe a very precise physical description will be required to be used in the prompts. All the models I have used have troubles correctly rendering a golf club, it was always drawn as bent, too short or strange looking, so at some point I gave up. This is the best result I got after ~50 tries:
If anyone is curious to try it by himself, here are the prompts I have used:
Positive prompts: (golf course), (flat chested), (long black hair), (petite), (twintails), (red eyes), (tight all black leather jacket), (tight all black leather trousers)
Negative prompts: hat
Model: Anything V4.5
CFG scale: default (10)
Denoise strength: default (.5)
I recommend to use it on morning time (UTC) during the week before 12.30, when itās much less crowded.
Varthall
I love the Kooh artwork that came out. Though her right index finger looks like it was glued in.
I guess the model was attempting to render her hand holding a club, judging also by the position of her arm. Her left ponytail also looks like itās midways shifted towards her left elbow. Unfortunately there were always errors like this on every render, though I have stumbled upon some prompts that might help fixing this. Iāll try adding them from now on, although Iām not sure if they have any impact:
Positive prompts: high detail, highly detailed, hyper realistic, intricate, intricate sharp details.
Negative prompts: lowres, bad hands, error, missing fingers, extra digit, fewer digits, cropped, low quality
EDIT: unfortunately they donāt, plus the āgolf courseā tag seems to be left out most of the times. Better to use fewer prompts I guess.
Varthall
Looks like someone is trying to generate new images using AI trained on SeeDarkās styleā¦ still needs some tweaking, but itās a starting point.
Varthall
Oh my god, I donāt believe it with my own eyes! This is incredible!
I was thinking I might try to train LoRAs for StableDiffusion based on Kooh, Hana, Spika etc and try to generate new art of them. I like combining AI with my own moderate artistic talent to create stuff.
I guess Iāll try making some LoRAs this week and then see if they play well with the SeeDark LoRA.
Iām really looking forward for your creations, please post them here once theyāll be ready. Unfortunately I donāt have neither a powerful HW nor artistic talent, but I have uploaded the model on SeaArt (https://www.seaart.ai), as soon as it will be reviewed Iāll start making experiments with it.
Varthall
So far Iāve worked on a Kooh LoRA. It gets her overall look (twintails, red eyes, tends to be black hair) right, but itās not quite right on the overall feel - and it easily blows out the style and looks funky. Iāll need to work on the dataset and training settings more. I foresee Hana and Spika and Lucia posing a greater challenge as thereās less to work withā¦
Keep it up man. I hope to see more Pangya character LoRAs in the future. Hopefully we can see them in different art styles.
Nice, keep up the good work! Iām wondering if the model includes tags for each character, so that you can use their name in the prompts without the need to describe them. Also, Iām afraid the model hasnāt been trained on Spika, since she was introduced 3 years after SeeDās passing. Still no news from SeaArt, the model is still under review, no idea when it will be ready for use. For us without a powerful HW setup, SeaArt unfortunately seems to be the only free service which offers the option to upload new models.
Varthall
