For me, I have spent more time than I want to know a number for, include lots of googling for artists and art techniques that might end up doing something when used as a prompt. It is much like learning a programming language.
Yeah, that too. Obviously most people went into it the same way they did an hour prior and got dramatically different results. They’ve already pulled the beta down (I think primarily because of the amount of nude smut being generated) for some retooling so many people’s options were formed early and without the benefit of time to learn how it can potentially operate.
What do you mean by “nude smut”, and why would they care about it? Do you mean pictures of nude models rendered with celebrities’ faces?
Nah, all sorts of nude depictions. I assume there’s potential for liability with the real faces thing (celeb or otherwise; I’ve seen people import faces via whatever magic), issues with Discord’s TOS, potential for child depictions, but I think the single largest factor is just not wanting MidJourney to be primarily a stream of dudes making pictures of naked women or having new users join and enter a channel to create and see that’s what the channel’s full of.
There’s already a word filter in place but there’s a nigh infinite number of creative ways to say “naked” despite the channel agreement when you join saying no adult content. And there’s debates on how/if to police it, segregate it, whatever but I’m not deeply involved in that or looking to defend either side.
An example: I once tried the prompt “Jesus family portrait | wanted poster | hyperdetailed colored pencil drawing | diorama | 8k resolution | hyperrealism | volumetric lighting | photorealistic | tijuana bible | junji ito | kei toume | tim burton” (I wanted weird, I got weird.) I tried the same prompt with the new system
Top row is classic “coherent”, bottom row is “stable”. The differences are very.
Ok. I take it back. I’m really loving the results I’m now getting with the Stable Diffusion. Like I said, you have to jam some modifiers into that one line or you’ll just get something weird - a stock photo crossed with a Photoshop fail. I like the fact that I can generate four images at once but then I get FOMO. I want to publish them all. Besides, a cluster of images when you’re perusing the feeds hurts my eyes.
That’s a might fine car wash!
I installed Stable Diffusion on my own PC to run locally, taking advantage of my RTX 3080Ti’s 12GB VRAM and now the man can’t tell me how many times I can run prompts. It only run at about 512x512 (can push further if I sacrifice returns per run but not by much) but it does knock out images surprisingly fast. Played around with the tilt-shift effect which was fun and neat
Thanks! But then I should be more specific about how I got it—asked for an “eldritch cheerleader washing a car” rather than an “eldritch cheerleader carwash”.
I tried running it on a 3080, a 2080, and a 1080 and that does not sound right. It’s true that with the default settings a lot of memory is used, but you can set n_samples to 1 (and also turn off the stupid watermarking and nudity filters that just take up memory without doing anything useful). Also, for really high-resolution attempts, you can use the “optimized” script which takes much longer but uses a lot less video memory.
One problem could be, not the memory consumption itself, but the fact that the model was trained at 512 x 512, so that weird artifacts start cropping up if you start increasing the dimensions. I have noticed some of that.
Poor phrasing on my part. I was saying that I have it set to return 4 images per run. I could enlarge the resolution but I’d have to decrease the images per run. Since I’m mainly just dorking around for funsies, the number of images matters more to me (immediate returns on my dumb idea prompts) than high resolution images that I’ll never use for anything practical.
I’m running it without filters, mainly as a matter of principle. I fully support a service filtering when they’re the intermediary in creating the image but not if it’s locally hosted and more akin to a pencil or Photoshop than a live service. Plus, at least on MJ, the word filter is pretty heavy and blocks a lot of common terms/phrases. No idea if that was the case for SD in general as well but I didn’t feel like finding out. And watermarks are dumb.
The new system can make such nice images overall it is a heartbreaker what a bad job it does with hands. On the other hand (…) there is one where it did such a bad job with hands that it made not only one of my favorite images from Night Cafe but one of my favorite images from Earth.
Another of my favorites from everything I’ve made on Night Cafe is this “robot yokai by Mark Ryden”.
Here is a sample of things made with the new system that I especially like. Some are Mark Rydens (with prompts like robot girl, robot yurei, robot yokai, robot Elvis, ghost, ghost parade, etc.). Others are from recycling two of my old complex prompts (some parts of which I’m sure I can jettison.)
The two complex prompts:
Summary
Yurei | tronie | wanted poster | hyperdetailed painting | diorama | 8k resolution | hyperrealism | volumetric lighting | photorealistic | golden hour | kei toume | dan witz | margaret keane
Yurei parade | full-length portrait | tronie | crackled oil painting | Dan witz | margaret keane | Zdzisław Beksiński | kei toume | junji ito | dungeons and dragons guidebook | postcard
(Facehands lady was generated by the second one.)
How much space does it take? (Just curious, my computer runs on old integrated graphics and isn’t remotely up to running something like that locally.)
It’s mainly GPU dependent. I think the minimum is supposed to be 8GB VRAM but some people have kludged it working with 6GB. I have 12GB and it’s using 11.83 of it when rendering four (two sets of two) 512x512 images with 150 steps each. I’m also up to 12.75GB of CPU system memory in use (though some stuff is running in background as well)
I mean how much hard drive space does the data set take.
My folder for it is 5.83GB, not counting whatever the Python client takes. Actually, about 800MB of that is rendered images I need to clean so closer to just 5GB.
In the coherent system I did “children trick-or-treating by Norman Rockwell”. The new system bans use of the word “children”. (But “trick-or-treating by Norman Rockwell” still gives results.)
Looks to be the same in Artistic which isn’t surprising (assume it’s just a site wide filter) but sort of pointless since you can’t make people-looking people in NC Artistic anyway. But, yeah, the filters are getting silly in how obstructive they are.
That’s kind of goofy. I tried to do a piece called, “My Childhood Home,” and it kept erroring out - I guess that’s why. But then why did it let me do a piece called, “Ok, Not My Childhood Home”?
I keep getting emails on how ‘Stable Diffusion’ will be available next week. I wonder if it will be much different from the beta version.