I’m in a “wait and see” on it, especially when it comes to non-photographic, artistic style renderers. I’ve never been especially impressed with Stable Diffusion in that regard and, although Loras exist and are awesome, it’s nice to have thousands of styles baked in.
But, as I’ve said before, I own and use Automatic1111 so I’ll definitely be trying it out. Just not ready to ride the hype-train.
It also doesn’t look like it can do the level of “weird” that I like. Here is the same prompt in 1.5, 2.1, and whatever point beta of XL is at Night Cafe:
XL is obviously the most aesthetic and realistic, but that wasn’t what wanted, I wanted random crazy stuff.
Yeah, that’s part of what I mean. Stable Diffusion doesn’t seem to stylize as well and, when all the hype examples are photorealistic, I’m skeptical that it’ll scratch that same itch. For the record, your prompt (unchanged) from MJ:
I guess there’s also still the big question of nudity. SD2.0 and 2.1 (?) stripped most/all of the nudity out of the model which really hurt people’s interest in it. Regardless of how someone feels about other people making fake nudie pics, the amount of new models and support for 2.0 is a shadow of what 1.5 still enjoys today. Like the old joke about all internet innovation being driven by porn, even if YOU don’t want to make those things (and obviously a lot of people do not have that interest), the tools you DO want to use were likely developed in part in the quest for a better boob. 2.0 has a number of other issues as well but few people seemed as interested in tackling or smoothing those if they couldn’t render the images they wanted from it, leaving everyone else without the tools and stuck with a flawed base model. It sounds as though they’ve been cagey on whether or not nudity is back in SDXL and that’ll probably make a significant difference in how widely its adopted and supported on model-hosting sites like Civitai.
Those are great. That’s what I would hope for in SDXL–more realistic, but still weird.
XL does have nudity. Clipdrop occasionally blurs output because of it. I obviously don’t know what the actual content of the output being blurred, but sometimes it slips through. Like this higly erotic unrequested nudity:
(You sometimes do get more boobie-related stuff that slips through, too.)
Well, “good” nudity and on-demand rather than accidental 
I’ll admit I haven’t been following SDXL very closely – more of a “I’ll play with it when it’s out” thing – but the responses in the Reddit thread you linked last night make it sound like it’s still an unknown.
Sometimes I like to expand a painter’s canvas to see what he actually saw while painting the piece. Here’s the view Claude Monet saw while he painted this Water Lilies piece (original painting within the white-stroked box).
The right side of that is really aukward.
Midjourney and Adobe Firefly work well together. Midjourney to generate characters, Firefly to merge them together, and Photoshop to do the touching up.
It’s always fun to take the kids to the circus. Just be careful which circus you go to.
…and don’t harass the clowns.
I didn’t think the mainstream sites were trained with a lot of hard core pornography, but try using the keywords, “Tonight I’m going to rock out with my cock out,” at Playground AI and prepare for eye-bleach if you can get a result.
You just need to rephrase it a bit. Try “Rooster with a guitar standing on a rock” in Dreamshaper.
Midjourney, Leonardo and especially Adobe Firefly certainly ban anything close to pornographic prompts. I don’t do porn, but Firefly wouldn’t even let me generate a bare foot, or work anywhere near a woman’s secondary sex characteristics to retouch. Although I think Adobe relaxed a tad with their latest update (lot’s of complaints, no doubt). And, yes, you need to be creative with your prompts and persevere to get the results you want.
Who says I didn’t get the results I wanted?
Over in a Facebook group people keep trying to find ways to trick Bing into making pornish content. Lots of fun with geoducks.
Night Cafe now has SDXL 0.9, but they are charging a fucking insane 6 credits per image. You do get a 10 point bonus for your first SDXL 0.9 published, though, so here is the one and only image I’ll ever make at that price:
(Prompt is Ugly cute scary funny weird horror calavera animal photograph | by Dan Witz, Mark Ryden, Margaret Keane, Pino Daeni, Carne Griffiths, Wadim Kashin | Polaroid, Kodachrome, Ektar | Professional photography, bokeh, natural lighting. This is what the prompt produces in SD 1.5.)
Night Cafe has also recently stopped awarding credits for posting to Twitter. That used to be a guaranteed 30 credits per day, more than enough to keep a piker happy. Not anymore. I just create a few pieces and post & vote in the challenges for a big 7 credits per day. Then there is the 5 day, 30 day, 100 day streaks, but still, they’re getting chincy over there.
I guess Nightcafe dropped the cost for SDXL to 3 credits? They sent me an email so I dropped by for the first time in ages and ran some recent MJ prompts through it just to get a feel. I should note some important caveats: The MJ images were cherry picked out of my popular recents and came out of multiple grids whereas the NC SDXL images were the first that popped up since it was still costing me credits and the NC model is less welcoming to just mucking around with rerolls. Also, I didn’t play with image dimensions/resolution in SDXL.
In short, this was just a “Wonder how SDXL is” and not a fair & balanced cage match meant to prove anything.
First image in each is Midjourney, second is SDXL 0.9 via NightCafe
Haunted eldritch church, eerie, inkwashes, ink pen, Aerochrome art
Redheaded fire efreeti woman dancing in the flame of a burning candle, feminine silhouette, translucent silk scarves of flame, shoulders, watercolor and ink on paper, dynamic lighting, editorial pose, loose flowing silk clothing, Art by Sophie Wilkins and Anne Dewailly and Aleksandra Kingo
Minimalist spirograph art, Ink art, Madonna, Mary, Grotto, by Ghada Amer and Harriet Lee - Merrion
Woman paladin in silver ornate armor, spear & shield, radiant, ankles, pulp fantasy, crayons and mixed-media, style of Rose Wylie
young woman wearing a gown of colorful cloisonne geometric glass which is lit from within, wearing glowing geometric colorful glass dress, Takashi Murakami inspired glowing stained-glass dress, low-shot, front facing, volumetric lighting, dramatic lighting
Wreck of the Edmund Fitzgerald by Edward Wadsworth and Raoul Dufy
Overall I’d say that I still prefer MJ’s results on balance but SD is closer than ever to a viable replacement and, again, I’m comparing these without spending real time on SDXL where I can learn its foibles and tweak my prompts more effectively. I’m actually pretty impressed with its handling of artistic styles since that was a weak point for it before and most of the “OMG LOOK” hype images I’ve seen for SDXL were photorealism so I was epxecting more of the same weakness.
You can use SDXL 0.9 for free at clipdrop.
And here is Bingdalle
Cool. I did a couple of others (didn’t change my opinion) but then it said I hit the cap unless I make an account so called it there. Not that making an account is great shakes, but my curiosity was already sated.
My Cat Benny, on the prowl:
Benny does ballet:
Clowns have feelings, too:
July 5th: The After-Party:
