Type | |
Stats | 187 |
Reviews | (3) |
Published | May 12, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 Strength: 1 |
Hash | AutoV2 2BC734FB02 |
Watercolor style for anime, one of my personal favorites. Use anylora model if you want the intended style, other anime models may give different results.
Its effects vary depending on your negative prompt, model, and CFG:
Without negative prompt of "worst quality" or "low quality", or similar embedding, it has a very strong watercolor style at strength 1.
If you do use that negative prompt, the effect is more subtle, so you can increase the strength a bit more to get the stronger effect (1.5 seems to be on the high end for this, you may want to try 1.3 though).
Tends to mangle faces at low resolution, use hires fix or ControlNet to get 1024x1024 images, or inpaint faces.
For a somewhat stronger effect without increasing the strength, but more painting-like, you can add the keywords "watercolor" or "watercolor painting" (some are shown in examples grid).
Technical details: Trained on a relatively large dataset with GPU for many, many steps and then fine-tuned on CPU with a huge batch size, which took forever, but was needed to increase its style while allowing it to generalize. It was pretty difficult to find the "sweet spot" in terms of training time, hypernetwork size, layer structure, etc. I tried larger ones or wider ones but they were worse and didn't generalize very well; smaller ones had too little effect. It's almost overtrained to the limit with low learning rate and dozens of LR restarts, but it finally had the proper watercolor style so I'm satisfied with how it turned out.
I'm trying to explore potential of hypernetworks.