https://civitai.com/models/1574800/dimsum-noobai-vpred

Read the contents below before use.

I have set a moderation on the gallery of this post, so the NSFW content may not appear on this page.

ABOUT THE MODEL

---

20250526v0.2

Trained on NoobAI-XL V-pred 1.0

The goal is to recreate Japanese and Korean-style girl. The current training is still in the early experimental stage, and there are still many issues.

The model will gradually increase the dataset size within a controlled scope.

Most of showcase images are generated at SDXL native resolution with ADetailer and Upscale with i2i sd sctript: SD upscale.

Updates

• The dataset was increased.

• Enhanced depth and lighting.

• Enhance realism and generalization capabilities.

• The caption method and training approach have been changed. caption is now generated using Joycaption2 to produce natural language descriptions, followed by manual refinement. The training method has shifted from DreamBooth to full fine-tuning.

Issues

• The teeth don’t look quite right.

• ADetailer is essential for the face; otherwise, it becomes blurry.

• The ‘from side’ angle doesn't perform well in full-body or cowboy shots.

If you encounter any issues while using the model, feel free to leave a comment.

USAGE GUIDE

---

Tips

• I recommend using SDXL's original resolution first and then pairing it with ADetailer to fix the face. And Upscale with i2i sd sctript: SD upscale.

• Tags like “realistic” can enhance the sense of realism.

Requirement

Recommended frontend: Forge

Sample : Euler

CFG : 3~5.5

Hires.fix: Recommended Upscale value is around 1.2~2.0, denoising strength is 0.1~0.2.

Recommended size: 768*1344, 832*1216, 896*1156, 1024*1024, 1344*768, 1216*8

32, 1156*896

If blurriness appears on the face, please use ADetailer to fix it. The recommended denoising strength is 0.1~0.4.

---

This model's license inherits from NoobAI-XL V-pred 1.0