Combining with dalle-mega produces some really cool results! (init image from dalle -> SD)

#5
by dlivitz - opened

If anyone is interested I can post more details, but basically taking the dalle-flow approach (manually), you can get really creative and detailed images.

Prompt "A royal crown made of green bok choy." (also works for just a crown, thought this was a little better, but it's not key)

Here I created this image with dalle-mega (picked from 9 images)
A_royal_crown_made_of_green_bok_choy_dalle-mega_S_135979_s_10_T_1_2022_08_15_18_47_38_8.png

Then fed it as an input image to SD, and got these 9: (strength 0.5, scale 12)
image.png

But if you just try SD, it doesn't really get it:
image.png

I wonder if anyone has any other tips/successes with combining these models?

I wonder if anyone has any other tips/successes with combining these models?

Thanks for sharing your results! I tried the same thing earlier after I installed the Stable Diffusion weights on a local machine yesterday (since the command line interface makes it easy to try this and had an image init example), but just as a one-off image quick experiment, rather than with a grid, and also had some success on prompts -> dallemini(mega) output, and then used those as init images for SD.

Stable Diffusion out-of-the-box isn't as good as DalleMini on some prompts. e.g. "GoPro footage of the Roman invasion of Britain" looks a lot better in DalleMini than SD (https://twitter.com/hardmaru/status/1558831284838797312)

But for some prompts, SD does better, or slightly better, out of the box:

“Competitive sleeping at the Olympic games.” https://twitter.com/hardmaru/status/1558977103915548672

"Pixar Animation of Mad Max Beyond Thunderdome" https://twitter.com/hardmaru/status/1559186273759563776

Using the method you described, this will be a nice technique in the toolbox!

Wow really cool thread here!

Sign up or log in to comment