Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Cardinal Cardinal]

Not the trend I was expecting.
View attachment 5502902

Needs more dogs.

 

[CervixHammer]

Needs more dogs.

There's always puppies in the lemonade stand.

 

[whatever I feel like]

I see your Frozen dogs and raise you some Scary Starbucks Spiders, which I am now making purposely because fuck it.

edit: Adding a few more, I do like the AI randomly deciding to make them black women in one of the images.




edit 2: I tried to replace Starbucks with Gaza, Palestine and Lybia in the prompt and they are all blocked. This is bullshit, Microsoft is censoring the plight of Palestinian spider monsters everywhere.

edit 3: Don't stick your dick in spider crazy.

 

[TapewormSalesman]

I see your Starbucks Spiders and give Burger King Buggers

 

[Sinner's Sandwich]

What did Bing mean by this?
View attachment 5499406View attachment 5499408

The frst two pictures could be a scene from a Batman movie showing Killer Croc in his human form turning into a monster.

 

[indomitable snowman]

I see your Starbucks Spiders and give Burger King Buggers

Those first 2 are almost kinda cute. The 2nd has a sort of 80s scifi prop for cricket people ascetic.

 

[Accidental Racism]

I see your Frozen dogs and raise you some Scary Starbucks Spiders, which I am now making purposely because fuck it.

edit: Adding a few more, I do like the AI randomly deciding to make them black women in one of the images.

View attachment 5503234View attachment 5503235View attachment 5503252View attachment 5503253View attachment 5503254View attachment 5503255


edit 2: I tried to replace Starbucks with Gaza, Palestine and Lybia in the prompt and they are all blocked. This is bullshit, Microsoft is censoring the plight of Palestinian spider monsters everywhere.

edit 3: Don't stick your dick in spider crazy.

View attachment 5503266

I hate being aware of the fact that somebody, somewhere, would crank it to these

 

[CervixHammer]

By the way, if you've noticed that Bing Image Creator now says it's part of Microsoft Designer, they've got an AI-assisted image editor. It's got a background remover, various filters and templates, icons and stock graphics, plus you can upload images (and fonts if you don't like its selection). As you muck around with something, it will start spitting out variations that you can click on and refine.

For example:

 

[Post Reply]

By the way, if you've noticed that Bing Image Creator now says it's part of Microsoft Designer, they've got an AI-assisted image editor. It's got a background remover, various filters and templates, icons and stock graphics, plus you can upload images (and fonts if you don't like its selection). As you muck around with something, it will start spitting out variations that you can click on and refine.

For example:
View attachment 5508710
View attachment 5508717
View attachment 5508730 View attachment 5508751

No wonder Patrick hates AI so much.

 

[Gourmet Race]

New AI animation just dropped, but before I show it, let's talk about AI animations during early 2023. This used to be peak AI animation:

Clothes, face and body were inconsistent, they had to use a normal background because it wasn't suitable and the animation was good for its time, but it looks rough nowadays. I forgot when this was created, but this video blew up somewhere on March 2023.

Now let's forward to April 14, 2023 when this video blew up on Twitter:

Nothing in this video is still consistent, but there's one improvement for sure and that's the animation, it's smooth as butter.

Now let's jump to the present, which is the 11th of November, and take a look at this:

Background is still inconsistent and the hair too (though it's not that noticeable), but the clothes, face and body don't change like it's a time-lapse anymore. The person who made this is cfsstudios, and he uploads stuff like these on TikTok. The improvement of AI animation in the past few months is astonishing. It went from constantly changing pixlr image per frame to this.

Anyway, here's another video they made, it's similar to this but day time; it's less consistent, but still pretty good:

 

[whatever I feel like]

Who's to say that AI won't let you eat babies? Enjoy these Taco Bell Ticks. And no, Bing does not, seemingly, know what a tick looks like.

 

[Puff]

I've not worked on SD since XL came out. Is there a way to make it run on more potato-y machines? I'm having out of memory errors at the last iteration every time. Or am I just doing something wrong?

 

[Baraadmirer]

That's the thing: I don't want to just completely redo one image into another. You know how there are LoRA's that add extra detail to what you're generating?
View attachment 5479462View attachment 5479463
I want something like that but having it work on a pre-existing, non-txt2img image. I want it to keep all of the original edges, shapes and colors and only add extra detail to it.

You're probably looking for inpainting. Under the img2img tab (if you're using AUTOMATIC1111) there's an "Inpaint" subtab just above where the reference image goes. You can navigate to that tab, place your reference image there, then mask over the parts of the image that you want to change. Note that at this point your prompt is what you want to appear in that masked region, and a high denoising strength like 0.8 is strongly recommended. Of course, you can use LoRAs in img2img, but you're going to have less control over what areas you want altered.

Currently in the middle of training my first LoRA on SD 1.5, so if anyone's got tips they're very much welcomed.

As image tax, I present the de-animeification of a JRPG character portrait. Prompting was very minimal and I used a photorealistic model for this. From

Spoiler: Original image

to

Spoiler: Generated image

I've not worked on SD since XL came out. Is there a way to make it run on more potato-y machines? I'm having out of memory errors at the last iteration every time. Or am I just doing something wrong?

I'm also chickening out on using SDXL cause I worry about my computer giving out halfway.

 

[inception_state]

I've not worked on SD since XL came out. Is there a way to make it run on more potato-y machines? I'm having out of memory errors at the last iteration every time. Or am I just doing something wrong?

What UI are you running? With A1111 try starting it with --medvram or if that doesn't work, --lowvram. If you have less than 8 GB of VRAM I wouldn't bother with SDXL though. It will take too long to generate images to be worthwhile even if you get it (barely) working.

 

[Puff]

What UI are you running? With A1111 try starting it with --medvram or if that doesn't work, --lowvram. If you have less than 8 GB of VRAM I wouldn't bother with SDXL though. It will take too long to generate images to be worthwhile even if you get it (barely) working.

I was assuming it was something weird since I didn't get an error until the last step in the generation (regardless of iteration number). Either that or some kind of settings issue. I'm hanging around at 10GB of VRAM so I'll give those settings a shot.

 

[whatever I feel like]

You know, fucking Bing chooses the one where they're god damn ticks to show me panty shots. I wasn't even asking for any.

 

[DavidS877]

I was assuming it was something weird since I didn't get an error until the last step in the generation (regardless of iteration number). Either that or some kind of settings issue. I'm hanging around at 10GB of VRAM so I'll give those settings a shot.

10GB is pretty not-potato. I switched to ComfyUI on my 8GB cards and have been happy with it. The learning curve is steep though.

 

[Baraadmirer]

10GB is pretty not-potato. I switched to ComfyUI on my 8GB cards and have been happy with it. The learning curve is steep though.

is 8GB alright for SDXL? I'm pretty comfortable with the wide variety of LoRAs and models for 1.5 right now.

 

[DavidS877]

is 8GB alright for SDXL? I'm pretty comfortable with the wide variety of LoRAs and models for 1.5 right now.

It's working fine for me. It tends to have to load and unload things as it processes so if you have slow drives it can get a little slow, but even with an upscale I've not had any issues. This is on Linux, the one concession I do make is not to do anything intensive on the system like running Chrome that chews up a bunch of VRAM while it processes. Since they're web UIs I tend to control from a system not doing any of the processing.

I'm doing 40 steps, and then 20 upscaled steps(2x I think) and a 3060ti 8GB is coming in at about 95 seconds.

 

[Puff]

10GB is pretty not-potato. I switched to ComfyUI on my 8GB cards and have been happy with it. The learning curve is steep though.

I hear it's a lot faster. I think I tried to install is a few months ago and my internet shit itself so I never got it going. Attempt two starting. And Yeah, it's good enough of a card, but not exactly state of the art. I'm going to be hitting a wall with AI stuff sooner rather than later.
Edit: Why the fuck does this not allow me to resume downloading if something goes the least bit wrong. FFS. Is there a trustworthy source that's not github? It REALLY doesn't like my connection


Edit edit: ComfyUI working after downloading on my phone and forcing a driver update. It's faster with XL than I was doing SD on automatic1111