Mondrian Variations 3 - SD
Mondrian Variations
If you work with both Midjourney and Stable Diffusion, then you know the strengths and weaknesses of both engines. Trying to compare the two is tricky, especially if you need to explain those differences to someone who hasn’t used either.
But there’s a key distinction that would be immediately obvious — asking them to interpret what “in the style of (artist)” mean. When you look at the Midjourney examples before, there is a very deep level of creativity, with results that come as a surprise. In contrast, asking Stable Diffusion for something similar and you get fairly expected results. This is not necessarily a bad thing, it’s just… different.
Workflow
For these Stable Diffusion sets, I used a variety of checkpoint models. Besides stylistic differences, I also tried different ones to see if perhaps one model simply lack the training data for Mondrian. I also used several different generation methods. The images generated used a combination of the following — sometimes one of them, sometimes multiples of them. I was mostly exploring what I could do to insert more Mondrian into the mix because the “out of the box” results from txt2img only were not very satisfying.
- txt2img. Using “By Mondrian” or “Mondrian” alone didn’t seem to have much of difference. Many of these do use a strong weight since it gradually got weaker when the prompt gets complicated. I used a weight around 1.4-1.6 for most of these images.
- Control Net: Canny. I supplied an actual Mondrian painting with black grid and color rectangles for the CN. This is the one whre it became the backgrounds. It does pick it up.
- Control Net: Reference Only.
- img2img. Supply actual painting by Mondrian and Denoise at 0.9 - 1.0. This created some of the more interesting results, especially when combined with CN Reference Only.
I do plan on annotating each image on my site when I have the time to do so. I have picked out about 120+ images showing both good and bad renders, and 25 for Instagram split into 3 sets. So it will take some time for me to thoroughly go through them.
Models
I tried out a few new ones and some old favorites: (Warning: First 4 links go to a CivitAI pages with NSFW image examples. Last one is also a CivitAI link, but it is a strictly SFW model so it does not contain NSFW examples.)
- Virile Fusion v2.0: photorealistic 3d render
- Virile Reality v1.0: photorealistic / photography
- Virile Motion v1.0: pixar 3d render
- Airfuck’s Wild Mix v1.0: photorealistic 3d render
- A-Zovya Photoreal v2.0: my current preferred photorealistic model
Virile and Airfuck’s were made specifically to create gay AI art so the men do look “hotter” right out of the box. aZovya is a general purpose photorealistic model for SFW gens. It renders better skin and lights, but at the expense of the men being less “gay-appealing” out of the box. Virile Motion is a new model made by the same guy behind Virile Fusion and Virile Motion, for creating renders that look like Pixar 3d animations.
Related:
- Mondrian Variations 1: Midjourney
- Mondrian Variations 2: Midjourney
- Mondrian Zoom: Midjourney
- Mondrian Variations 3: Stable Diffusion
Images
These first three sets are assorted. I tried to mix them up a bit for Instagram posting because then there are variety. The 4th set, which contains around 120 images that are in order of generation, grouped by Stable Diffusion models and styles. Those are the ones I will be annotating in the future.
