Costume available for Crystal Shop.
Screenshots of the game's dress-up mode were used for training.
Description
Base Model: waiNSFWIllustrious_v150
Compressed by resize_lora
FAQ
Comments (4)
wow, can you tell me how do use the game screenshots and make it into datasets of the style of yours?
Sure! Here's my workflow:
1. Take screenshots in Project SEKAI's Dress-up Mode with the outfit you want to train.
Costume-only LoRAs: 20 shots (4 shots × 5 angles with multiple characters).
Also training hairstyle or unique accessories: 40 shots (4×5 with the specific character + 4×5 with multiple characters).
2. Crop each image to 1024×1024 centered on the character.
3. Upscale to 2048×2048 using ScuNET. (I don't recommend RealESRGAN_x4Plus Anime 6B here — it tends to hurt background removal accuracy later.)
4. Remove the background with BiRefNet. This is basically required if you're training on Noob or Anima base models. For IllustriousXL and similar, you might be able to skip it.
5. Manual touch-ups — this step matters more than it sounds!
6. Downscale back to 1024×1024.
7. Tag and train.
Note: If you skip the manual corrections in step 5, you can also skip steps 3 and 6 — but in that case I'd recommend increasing your image count by about 1.5–2× to compensate.
@NekonoMikeko I have a question. In the game, some NPCs like Elena have only live2d images but no models. How did you use which AI generator to complete the rest of their bodies? And the style is maintained consistently as well.
@nepneko Hi there, I'm still in the process of trying different things, but for the past month or so, here's how I do it:
- Use the original image as a reference in Nano Banana to generate a full-body three-view sheet. Any parts not visible in the original image get specified at this stage. If something isn't shown in the illustration but appears in an official Illustration, MV or similar, I match that — otherwise I just make it up. For Elena, for example, I went with "white short shorts, bare legs, no socks, white and green sneakers."
- The three-view sheet is really important, so I redo it until I'm happy with it. Nano Banana can be kind of dumb and lazy, so I often end up doing manual corrections by hand and then cleaning it up with Stable Diffusion's img2img.
- Once the three-view sheet is done, I use it as a reference in Nano Banana to generate 20–30 training images with varied poses and compositions.
- Again, Nano Banana is kind of dumb, so manual touch-ups are pretty much a given.
- Remove background, tag, train.
If you're only generating a few images, I think you could probably go straight from the three-view sheet without making LoRA. And if you add a few illustrative images, you might not even need the three-view sheet at all.
The LoRAs in the past—Kotaki Nagi, for example—were trained directly on a game image and do not contain any information about lower-body outfits. As a result, the styles are likely to vary considerably.