v4.0 Changelog

v4.0 is a checkpoint trained on thousands of images from popular and recent gacha games, primarily miHoYo titles. A100 GPUs were rented for training, as local hardware would take years.
After training, the model was further trained to remove low-quality concepts. By default, outputs are higher quality than without quality tags, and even more detailed with them.
Finally, a simple merge of my JANKU-style LoRA (trained on previous JANKU images) was applied. Merging too much risked losing trained knowledge, so only a small merge was possible.
Compared to v3.0, images are way more detailed (see below on how to adjust). A shift towards anime style while keeping the amazing lighting shades from v3.0.

If you want less details or issues with hands, lower CFG to 3 and/or try scheduler Simple instead of Normal. Onsite gen works better with CFG 3, no clue why. If you use my positive embeddings, get the new lazypos and lazynegfrom the page. lazypos v1.0 is not good for v4.0. lazyneg v3 is optimized for JANKU. v4.0 is so good, you don't really need positive or negative prompts to start with, try "1girl" prompt and nothing else yourself and see the magic! 🔮

👉 Join me on our discord server (https://discord.gg/tWrKk3jcXx)! Share art, ask for help, or just chat. We have CivChan bot plus games there!

✨ JANKU v4 NSFW Trained + NoobAI EPS + RouWei Illustrious XL ✨