🎨✨ Dream Brush ✨🎨
Bring the magic of classic animation to life!
Dream Brush is a custom Stable Diffusion checkpoint lovingly crafted to capture the timeless charm of golden-age animation. 🌟 Inspired by the lush, hand-painted backgrounds, soft color palettes, and whimsical character designs of vintage animated films, this model transforms your prompts into warm, nostalgic masterpieces. 🖌️🐦
Whether you're dreaming up fairy tales, charming animal companions, or storybook adventures, Dream Brush delivers that signature magic touch you've been looking for. ✨🏰
📜 Features:
Soft, painterly shading 🖌️
Vibrant but gentle colors 🌈
Expressive, classic-style characters 🎭
Fairy-tale backgrounds and cozy atmospheres 🌳🏡
Ideal for fantasy, storybook, and magical settings ✨📚
Description
FAQ
Comments (5)
Great work!
Thank you very much, glad you enjoy it!
TL;DR: It looks like the dataset is quite small. It depicts almost the same things with ambiguous prompt. I think it's overtrained.
There is a problem with generating. I mean, there a a LOT of bad artifacts when generating even simple things. Especially eyes. Of course, hiresfix is extremely helpful, but without it the future of Dream Brush looks gloomy and messy.
I tried different schedulers, but even default "Euler" and "Euler a" are no good. These ones should work everywhere. But It just needs hiresfix or adetailer.
How did you train this checkpoint? What is the base model? How big is dataset? For me it looks like it's somehow undertagged (autotaggers are bad without editing, you know) and small. It loves to draw clouds with the same shape and the same composition. It clearly leans towards depicting blonde girls in red clothes.
Overall style is good, but lack of tagging and small dataset show itself.
I think you should make bigger dataset, add more good eyes pictures and partially retag it.
Thanks for your feedback. Can you supply me with some of the prompts you are using? Im trying to replicate the issues you are having. By training standards, the dataset was considered medium/large. And before I release the model, i test with a single word prompt with no hires and no adetailer. If it produces good images, I start again with a more detailed prompt and the use of hires and adetailer. I want this model to be up to snuff with other models I have released.
@King_Hatchet BTW it was for 1.0, prompt was just standard
quality tags, 1girl, outdoors,
Yep, this is it