Winged Cloud (Sakura Dungeon)

I am admittedly kind of a noob at this, but if you enjoy my humble work, any donations are greatly appreciated: https://ko-fi.com/insult_to_ninjary

Your generosity helps me stay motivated to keep improving and to keep trying new things!

Trigger Words:

Main trigger: winclosakdun

Helper prompts: parody, style parody

• The LoRA can potentially work with no triggers, but it may not be strong/consistent.

• For a stronger effect, use the main trigger: winclosakdun. You can also weight it numerically or with parentheses - I usually do (((winclosakdun))).

• If the (weighted) main trigger is still not enough by itself, next try one or both of the helper prompts along with the main trigger. They can be weighted as well - I usually do them all together, like (((winclosakdun, parody, style parody))).

• Please note that the main trigger here is only for the Illustrious model - it won't work for the SD1.5 models. The helper prompts may or may not work for SD1.5 (I never tried them), but they are not necessary as both versions of the SD1.5 LoRA work just fine with no triggers!

At a Glance (Illustrious):

• Base model: Anime style (I use WAI-illustrious-SDXL)

• LoRA weight: 1 (but slowly go lower or higher if necessary)

• Tested parameters: Euler a, Automatic, 30 steps, CFG 5 (but feel free to experiment!)

• Hires. fix: use a quality upscaler (like 4x-UltraSharpV2); denoising strength 0.5 - 0.3

• Extra: use ADetailer to fix faces and/or hands if necessary

Illustrious Version Description:

This LoRA produces art in the style of some of the Winged Cloud visual novel games, and specifically it's based on the game Sakura Dungeon. I had a very hard time training the SD1.5 versions of this LoRA, so it was nice that the Illustrious version came out ok on the first try. It's not without its shortcomings though.

Notably, the LoRA is not great at fine detail, and it's especially pretty bad at hands. You will most likely want to use ADetailer to improve faces and/or hands on most gens. I didn't use it in the showcase images just to show you what the LoRA can do on its own, but the showcase images are also pretty cherry-picked, so you might need to be a bit patient to get acceptable images (especially if you're picky like me). It needs more testing, but adding "game cg" to the prompt has yielded some pretty nice results for me so far, so that's also something you can try!

For more potentially relevant information, you can check out the original SD1.5 descriptions immediately below. Below those, there's a troubleshooting section in case you're having any problems.

SD1.5 Version 2 Description:

v2.0 is trained on the exact same image set as v1.0, but with different training parameters that seem to make the LoRA work a bit more reliably. You should be able to use the LoRA more normally now, without a lot of the extra hassle that v1.0 needed. Although it's better, it's still not great at hands unfortunately. Par for the course for a lot of SD1.5 stuff... Clip skip 2 seems to help somewhat. I'm leaving the old description for v1.0 below. Some of that information can still apply to v2.0 if you're having trouble and need some tips.

SD1.5 Version 1 Description:

This LoRA will generate images in the art style of some Winged Cloud games, particularly Sakura Dungeon.

I really love Winged Cloud's library of games/visual novels, and I especially like their recognizable art style. While this LoRA was not trained on specific characters, hopefully it will allow you to recreate this lovely art style with characters and situations from your own imagination.

Sadly, I've had a hard time getting this LoRA to work reliably. I still wanted to put it out there though, as I think it's a good concept, and maybe you all will have better luck than me once you experiment with it a bit.

Here are some tips that I've found helpful. For best results, generate at 512x512 and HiRes fix up to 1024x1024. It seems to work better with square images. If you generate an image in landscape orientation, you'll get some really cool poses and compositions that look a lot like in-game CGs, but the anatomy will almost always be messed up. If you don't mind running a bunch of gens and then spending the time to clean up the good ones, you might end up with some really nice stuff, but I didn't have the patience for that, lol

If you want to stay true to the art style, use the suggested models for this LoRA: Doll Like Anime or Koji. I tried out 70+ different models with this LoRA, and those were the best. If you just want to make cool-looking stuff with a little Winged Cloud influence and don't care about straying from the source material a bit, then I also tried and liked Cardology Mix, Disney Pixar Cartoon type B, epiC2.5D, and EthernalDark Ruby (which I sadly can't find anymore), among others. [Way later edit: My Greatest Hits merge also works really well, and would probably be my go-to nowadays for the SD1.5 versions.]

Troubleshooting:

If you're having persistent problems with anatomy, fine detail, and/or other jank, just try your best with all the usual fixes - different prompting (including negative prompt), try a different base model, use a different VAE, try different clip skip values, adjust the LoRA weight, use adetailer, inpaint, etc.

Outro:

If you make anything awesome, please consider posting them to the gallery! That goes for all of my models! I really like seeing what kind of stuff you all are able to make!

If you have any questions, feel free to ask! I will try my best to help! Also, if you have any ideas for future models, please don't hesitate to share them with me!