Crew
Draped in their blue and white uniforms, the Yale College crew team basks in the summer sun, their muscles relaxing after intense rowing. They share light-hearted banter and hearty laughter, their camaraderie shining as bright as their perspiration-drenched foreheads.
Rendered as hyperreal oil paintings in the style of J. C. Leyendecker, using Stable Diffusion and SDXL.
Recently, @mid_bi_west posted a couple of paintings depicting men from the crew team. Those couple of paintings reminded me of some of my friends from my college days, as well as memories I had with some of them. I thought that I’ll make some paintings of my own, with Leyendecker as the inspiration.
I have never mentioned this before, but many of you have noticed something distinctive in my Midjourney painting. There’s a reason behind that. No matter what mix of artists’ styles I incorporate into Midjourney, there’s always a touch of J. C. Leyendecker in there. Even a tiny smidge, like 0.1, can give the figures that taller, leaner look I’m so drawn to. The kicker? I only found out recently from a friend that he was gay. I’ve always just been a fan of those sleek bodies he brought to life in his illustrations.
Maybe this clears up the big question about why so many of my Midjourney pieces have that vintage, yellowish tint. You see, all the Leyendecker works that were likely used as training data for Midjourney were old, hence the golden tones. Even though I was really just after his signature lean figures, his retro color palette also found its way into my work. Of course, I’d take a stab at tweaking the colors in Lightroom, but let’s be real — it’s not always a slam dunk.
Text rendering with SDXL
Yesterday, @mid_bi_west shared a painting of a guy wearing a Yale sweatshirt on UD. It came out beautifully, and I wondered if SDXL has improved in text rendering quality.
So I tried to render the word Yale as well as the capital Y in all of these paintings, and to my surprise in many of my renders — which also include a broader series with other art styles and artists — the text actually came out quite clean. Some of the letters did decide to play hide and seek, but most of them are quite legible.
Some of the Ys did end up resembling the New York Yankees logo. However, considering that Yalies are often seen donning Yankees baseball caps, these AI mishaps inadvertently resulted in a surprisingly accurate representation.
You’ll be able to see them inside the three sets on Instagram, and in the full line-up of 51 paintings on my site.
SDXL with Automatic1111
I made these renders using Stable Diffusion and SDXL. Previously, I had thought that it would be impossible to render with SDXL using Automatic1111 without much complications, but after further research, it appears that the workflow can be greatly simplified due to some additional extensions.
I don’t want to bore you with the technical details, but if you’re interested in how to render SDXL with Automatic1111, I’ve written a guide on my Docs site, together with my recommended settings and extensions that you should install.
I will update the docs when I get more experienced with SDXL. All of these paintings published today were rendered with the settings I have described in the guide.
I also took the opportunity to test out some new beta SDXL model, so these paintings were not rendered with the Base SDXL model, but rather the RunDiffusionXL model which is currently in beta.
In terms of technical parameters:
- I didn’t use hires fix. The renders from SDXL came out pretty nicely, and when I used hires fix previously, the quality didn’t improve dramatically.
- I didn’t use ADetailer or Face Editor to fix faces. When I tried to do that in the past, they had made some of the images worse, even though some of them did improve.
- My recommendation is that if you do decide to use these plugins to automatically in-paint faces, you should always select the option to save the before-fix image when working with SDXL.
Text prompt in Stable Diffusion with SDXL, with the RunDiffusionXL Beta model.
- SDXL Base: RunDiffusionXL Beta
- 50 steps, DPM++ SDE Karras, CFG 7
- 1024x1024
- SDXL Refiner: SDXL Refiner 1.0, 10 seps (20%)
- Post Upscale: Topaz Gigapixel AI HQ 4x: 4096x4096
- Post: Adobe Lightroom color correction
Images
Images
Images
Images
Images
Images
