AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

hydroceramics: Jan 8, 2014

RIP Syndrome posted:

Just guessing since I've barely tried SDXL myself, but it's possible a1111 makes bad assumptions for it if you upscale directly in the txt2img UI. If you're not using latent upscaling anyway, you might as well send to extras and do the upscaling there once you get an 1x image you like. That works for me (tried with 4x-UltraSharp).

SDXL seems like a big improvement. I can get stuff like this now with basically no effort:

There's still a bit of jank there, but the refiner can be convinced to turn the typical unidentifiable clutter into details that almost make sense.

Didn't have any luck replicating the style, but still had some real bangers anyway:

# ? Jul 31, 2023 21:28

Adbot: ADBOT LOVES YOU

# ? May 29, 2024 17:03

Sedgr: Sep 16, 2007; Neat!

Those look great!

# ? Jul 31, 2023 21:59

RIP Syndrome: Feb 24, 2016

hydroceramics posted:

Didn't have any luck replicating the style, but still had some real bangers anyway:

I used the DreamshaperXL10 alpha2 model from CivitAI with a mashup of style prompts I'd stored. img2img overwrites the metadata from the initial txt2img, but I think "retro anime, pastels, by moebius" should get you most of the way there. If it doesn't come out Moebius enough you can insist on 2d, flat cel shading, etc.

quote:

Wow. Are you using the vanilla SDXL model? I tried doing cityscapes with 1.5 models, but couldn't touch that level of detail at all. Ended up chasing flatter styles.

Sedgr posted:

Fan of this, it looks almost like a photo/image collage, and the light is great.

# ? Jul 31, 2023 23:11

WhiteHowler: Apr 3, 2001; I'M HUGE!

For people using SDXL under Automatic1111, what settings are you using for generation?

SDXL seems to be very finicky about resolutions; I read that it wants 1024x1024 or something with the same number of pixels? What resolutions have you had luck with? Is HiResFix still required/useful for non-square output sizes?

And I remember reading a while back about the newer models requiring far fewer rendering steps. Until this week I'd still been using SD 1.5 with Euler A and 20-40 steps. Is there guidance on what to use with SDXL?

# ? Jul 31, 2023 23:43

Sedgr: Sep 16, 2007; Neat!

RIP Syndrome posted:

Fan of this, it looks almost like a photo/image collage, and the light is great.

Thanks! Prompt was (wildflowers), Baphomet, mountain climbing, (cyberpunk), diffuse lighting, gothic, detailed, intricate, geometric, elegant, highly detailed, lifelike, photorealistic, digital painting, artstation, illustration, concept art, smooth, sharp focus, sci-fi, hyper realistic, oversaturate, dramatic lighting, colorful

1024x1024 is a good spot. I find 800x1000 works fine. 512x512 comes out like this.

prompt "something super cool" in SDXL at 5, 10 and 25 steps. I've generally stuck with 25.

5 step

10 step

25 step

# ? Aug 1, 2023 00:00

KakerMix: Apr 8, 2004; 8.2 M.P.G.

WhiteHowler posted:

For people using SDXL under Automatic1111, what settings are you using for generation?

SDXL seems to be very finicky about resolutions; I read that it wants 1024x1024 or something with the same number of pixels? What resolutions have you had luck with? Is HiResFix still required/useful for non-square output sizes?

And I remember reading a while back about the newer models requiring far fewer rendering steps. Until this week I'd still been using SD 1.5 with Euler A and 20-40 steps. Is there guidance on what to use with SDXL?

I saw somewhere the ratio chart that SEEMS to work, though I'll typically get something then render out resolutions in an editing program so I'm atypical anyway. Here is the ratio chart for Stable Diffusion XL 1.0:

640 x 1536
768 x 1344
832 x 1216
896 x 1152
1024 x 1024
1152 x 896
1216 x 832
1344 x 768
1536 x 640

# ? Aug 1, 2023 00:23

Swagman: Jun 10, 2003; Yes...all was once again peaceful in River City.

# ? Aug 1, 2023 00:49

Sedgr: Sep 16, 2007; Neat!

Swagman posted:

# ? Aug 1, 2023 00:59

hydroceramics: Jan 8, 2014

RIP Syndrome posted:

I used the DreamshaperXL10 alpha2 model from CivitAI with a mashup of style prompts I'd stored. img2img overwrites the metadata from the initial txt2img, but I think "retro anime, pastels, by moebius" should get you most of the way there. If it doesn't come out Moebius enough you can insist on 2d, flat cel shading, etc.

Wow. Are you using the vanilla SDXL model? I tried doing cityscapes with 1.5 models, but couldn't touch that level of detail at all. Ended up chasing flatter styles.

Fan of this, it looks almost like a photo/image collage, and the light is great.

Just using midjourney 5.2. Don't have the time/computing power to do anything else well.

# ? Aug 1, 2023 01:24

Sedgr: Sep 16, 2007; Neat!

# ? Aug 1, 2023 02:02

Hadlock: Nov 9, 2004

Swagman posted:

https://www.youtube.com/watch?v=Ok7GVnNP_Fg

# ? Aug 1, 2023 06:19

Clarence: May 3, 2012

More beginner questions.

How do I go about getting bigger outputs? (There has been a bit of talk in the last couple of pages but it goes right over my head).

I'm using Stable Diffusion. 1.5 maybe? Just checked - it's A1111

The maximum size I can generate is 512 by 768 otherwise I get vram errors (6 gig card). Was there a start-up setting mentioned a while ago that would allow slower generation to avoid the errors?

Failing that, or anything else within SD itself, what's a recommended upscaling tool(?) to try, ideally with a tutorial available?

Here's my launcher settings, for what it's worth.

Clarence fucked around with this message at 09:08 on Aug 1, 2023

# ? Aug 1, 2023 08:56

Sedgr: Sep 16, 2007; Neat!

I was going to suggest the low vram setting but looks like you already got it. Google around a bit for your specific video card and auto1111 and you might find a solution, everything moves like lightning so its a bit tough to nail anything specific down. If it errors out a bunch make sure you reset everything to default and relaunch from the batch file. I had a few instances where trying out different settings broke the whole thing until a restart, so it might not be the resolution specifically rather than a combination of things. On my 6GB card I was doing 800x600 and sometimes larger so it should be possible.

Have a goat that ate the special flowers. He's very very chill.

# ? Aug 1, 2023 13:58

RIP Syndrome: Feb 24, 2016

Clarence posted:

More beginner questions.

How do I go about getting bigger outputs? (There has been a bit of talk in the last couple of pages but it goes right over my head).

I'm using Stable Diffusion. 1.5 maybe? Just checked - it's A1111

The maximum size I can generate is 512 by 768 otherwise I get vram errors (6 gig card). Was there a start-up setting mentioned a while ago that would allow slower generation to avoid the errors?

Failing that, or anything else within SD itself, what's a recommended upscaling tool(?) to try, ideally with a tutorial available?

Here's my launcher settings, for what it's worth.

It's hard to get big images that also look nice, even when there's enough vram. SD makes the nicest images when you generate at the size it's trained with (for SD 1.5 that's 512x512. For SDXL it's 1024x1024). You can go higher, but the higher you go the more confused the output will be. It'll show up as repeated subjects, weird anatomy, people (or just y'know, body parts) floating in the air and so on. If you're making something that's inherently repetitive, like machinery or isometric cityscapes, you can make it work for you, though.

So the "proper" way to do it is to generate at 512x512 or close to it, and then use an upscaler. You can do this directly by enabling the Hires.fix checkbox. There's a special kind of upscaler called "latent" that will embellish the image with more details, but it tends to change the style of the whole image (often for the worse). You can get around it by reducing the "denoise" factor, but if you reduce it too much it'll leave noise in the image. Usually you can walk it down to 0.51 or so before things go bad. When you see something you like, it's a good idea to save the seed (hit the ♻️ icon next to the seed number) and do repeat runs to experiment with CFG, denoise factor and upscalers. When doing latent upscaling it's easier to preserve the image style when upscaling by less. I use 2x at most, but you can go lower, to 1.5 for instance.

Finally there are upscalers that don't add any details but are mostly about leaving smooth edges and preserving the micro-style of the image. You can use those in Hires.fix, but there's not really any point, since you can "send to extras" and do the upscaling there when you're done with the hard part. I think these upscalers are pretty lightweight. Out of the stock ones I've gotten good performance out of SwinIR, but there's a whole ton of them to try. Download more here and place in the models\ESRGAN\ directory. I don't really know which ones are good, but 4x-UltraSharp seems ok. There are many specialized ones like, for instance, 2x_Loyaldk-SuperPony which is trained to upscale MLP episodes only. Have fun.

To sum up the process I've used:

* Start small, increase until you hit vram/quality limits. Heavyweight step.
* Enable the latent upscaler, don't go too hard with it (1.5x-2.0x and as low a denoise as you can get away with). Regenerate.
* Send to extras and do a basic upscale with SwinIR or whatever looks best (2x-4x). This should be lightweight.

That should get you into the 2048x2048 range at least. Though I haven't tested this on 6 gigs of vram, so there may be hard limits I don't know about. Maybe try some of the official optimizations. --lowvram and --xformers seem like the obvious ones.

There are also more complicated ways to get more detail and higher resolution by taking an initial image and upscaling/embellishing it in parts. Manual in/outpainting is one way, and there are promising-looking extensions like this one that automate the process. That one doesn't seem to work with the most recent auto1111, though. It seems popular, so maybe it'll get fixed.

You can also download different models/checkpoints from CivitAI. Some are better at details than others. I've gotten good results from Dreamshaper and Juggernaut.

# ? Aug 1, 2023 14:21

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

If you lock the seed you can usually also use tile diffusion with 1.5 (i don't think it worked with 2.x so probably doesn't work with XL). This will let you add more details, You'll need to play with the settings and sometimes some images just don't want to upscale well with tiles and *must* make a mess these I just normal upscale. You do need to play with the tiled upscales like a normal image gen a bit to get good results.

I've posted this before but this was made without leaving A1111 in 1.5 and tile diffusion. I did some sub regions to vary the prompt in different places and you can see the hard lines for them if you look for them, that is not an effect of tiling unless you do no overlap. This was done on a 3070 Mobile with 8GB VRAM. (that final 3k upscale was a good 90 minutes though!)

# ? Aug 1, 2023 14:28

RIP Syndrome: Feb 24, 2016

Sometimes there're also system tasks competing for GPU resources. Once when I hit Win+Shift+Ctrl+B after a GPU crash, some nvidia display popped up showing a task list with some stuff, like the thing that thumbnails your images and sends them to Microsoft or whatever. Dunno how much vram it uses at peak, but I killed it and moved its exe to a quarantine zone.

# ? Aug 1, 2023 14:40

RIP Syndrome: Feb 24, 2016

Clarence posted:

By the way, I just noticed it says "Low VRAM (medvram)" and it's turned into --medvram. Try unchecking that and instead adding --lowvram to the additional options.

Post your spookiest posting stations!

# ? Aug 1, 2023 18:57

Count Roland: Oct 6, 2013

Swagman posted:

I was thinking about Bioshock recently and this image is a great fit that that creepy vibe.

# ? Aug 1, 2023 19:04

Clarence: May 3, 2012

RIP Syndrome posted:

By the way, I just noticed it says "Low VRAM (medvram)" and it's turned into --medvram. Try unchecking that and instead adding --lowvram to the additional options.

I think that's done the trick. Before if I checked hires.fix it would error quite quickly. This time I've tried a 768 x 512 with --lowvram and it's gone all the way through.

Thanks for the help everyone!

# ? Aug 1, 2023 19:57

Hadlock: Nov 9, 2004

Hadlock posted:

Yeah well for one, it's possible to compile a binary from python code so you don't need to install python, python on windows is getting better but it's still a pain in the dick. The distributable copy of automatic1111 should be a self contained binary, period

And by default it should be CPU only, which is slower (by about double, but whatever, if you can get it in the hands of 10x more people, it's worth it) but cuts out the cuda dependency entirely. If Nvidia wants to help streamline the install of cuda they can get their butt in gear and help add in a silent cuda installer

Add in an automatic model downloader (make it use whatever credentials are necessary, I guess) and now you've got a pretty reasonable, low effort program

Finally id name it literally anything besides the entirely forgettable automatic1111

ComfyUI is fantastic and has a --cpu flag I got it working in about 2 minutes on my no GPU poverty grade 8gb ram thinkpad. Was like 260 s/itr but I was able to generate a good image from it :v:

https://github.com/comfyanonymous/ComfyUI

code:

$ git clone git@github.com:comfyanonymous/ComfyUI.git
$ cd ComfyUI/
$ pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 xformers
$ pip install -r requirements.txt

$ python3 main.py --cpu

Total VRAM 7627 MB, total RAM 7627 MB
xformers version: 0.0.20
Set vram state to: DISABLED
Device: cpu
Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
model_type EPS
adm 2816
making attention of type 'vanilla' with 512 in_channels
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla' with 512 in_channels
left over keys: dict_keys(['cond_stage_model.clip_l.transformer.text_model.embeddings.position_ids'])
0%| | 0/20 [00:00<?, ?it/s]
5%|████▊ | 1/20 [04:24<1:23:35, 263.99s/it]
25%|███████████████████████▊ | 5/20 [21:46<1:05:04, 260.29s/it]
50%|████████████████████████████████████████████████ | 10/20 [43:03<42:49, 256.92s/it]
95%|█████████████████████████████████████████████████████████████████████████████████████████▎ | 19/20 [2:45:33<05:46, 346.28s/it]
100%|██████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [2:53:53<00:00, 521.66s/it]
Prompt executed in 5586.78 seconds
^C
Stopped server

Hadlock fucked around with this message at 21:07 on Aug 1, 2023

# ? Aug 1, 2023 21:03

KakerMix: Apr 8, 2004; 8.2 M.P.G.

RIP Syndrome posted:

By the way, I just noticed it says "Low VRAM (medvram)" and it's turned into --medvram. Try unchecking that and instead adding --lowvram to the additional options.

Post your spookiest posting stations!

zoom calls :argh:

# ? Aug 1, 2023 21:54

RIP Syndrome: Feb 24, 2016

KakerMix posted:

zoom calls

Don't dox my boss, haha... er hang on, someone's at the door

# ? Aug 1, 2023 23:59

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

angry Donald Trump as a baseball pitcher throwing a hamburger like a ball, ((television in the background showing news of "criminal indictment")), ((ketchup stains on the walls)), in Mar-A-Lago, realistic

angry and crying Donald Trump holding a hamburger as a baseball pitcher, (throwing the hamburger like a baseball) at a wall in Mar-A-Lago, (ketchup stains on walls), screen in the background with ((news program)) ((ticker text "criminal indictment")), realistic

# ? Aug 2, 2023 04:05

Hadlock: Nov 9, 2004

Swagman posted:

This is possibly one of the best images posted in this thread, fantastic :golfclap:

# ? Aug 2, 2023 08:41

KinkyJohn: Sep 19, 2002

Which 1.5 models have been trained to handle higher resolutions? I know wyvernmix can do 768x1024

# ? Aug 2, 2023 09:01

RIP Syndrome: Feb 24, 2016

Mercury_Storm posted:

I like the one that says WS CRIIMMAL INDUDUICENT UP! with the burger-shaped head crab.

# ? Aug 2, 2023 13:20

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

I did some messing around with comfyUI and SDXL on my laptop last night was very easy to setup. Is the base 1.0_0.9VAE just the 1.0 with a baked in VAE for 0.9? it's huge for a VAE!

I think I got comfy UI running in colab too (it boots, waiting on checkpoints to upload)

# ? Aug 2, 2023 16:03

RIP Syndrome: Feb 24, 2016

pixaal posted:

I did some messing around with comfyUI and SDXL on my laptop last night was very easy to setup. Is the base 1.0_0.9VAE just the 1.0 with a baked in VAE for 0.9? it's huge for a VAE!

Yes, I'm pretty sure it's the whole thing including the VAE. I think they reverted to an older VAE because the 1.0 one had some bad artifacting.

GH issue: https://github.com/huggingface/diffusers/issues/4310

RIP Syndrome fucked around with this message at 17:15 on Aug 2, 2023

# ? Aug 2, 2023 16:28

AARD VARKMAN: May 17, 1993

# ? Aug 2, 2023 17:51

RIP Syndrome: Feb 24, 2016

AARD VARKMAN posted:

That chair's gonna give you a bad back.

I tried inpainting the hood area and sort of hosed it up. Can't find a surefire way to do it in a1111 without leaving halos.

# ? Aug 2, 2023 18:35

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

I'm playing with SDXL 1.0 with the 1.0 VAE because that's what I was uploading earlier and then had to troubleshoot some errors

did some known figure and upscaler testing figuring out the Comfy UI I layout I found
this one is something I've seen from no model, Trump was given facial hair (and it's even unprompted unless tired and haggard count)

This handles prompts a bit differently need to figure out why his hand is in 4 pieces

Based off the Danny prompt (it still includes his name horrifyingly enough)

okay I think I needed more supporting terms it's figuring it out

It really keeps trying to go back to this comicbook look, not that I don't like it but I don't always want it.

maybe someone has an idea

I'm sure it's in the mess of settings still at default

pixaal fucked around with this message at 18:53 on Aug 2, 2023

# ? Aug 2, 2023 18:51

RIP Syndrome: Feb 24, 2016

pixaal posted:

It really keeps trying to go back to this comicbook look, not that I don't like it but I don't always want it.

It usually helps to specify the style you want, e.g. "photo". Otherwise you get whatever. You can be more specific about the imaging tech, for instance try "35mm", "kodachrome", camera models and so on. E.g. "hdr polaroid" (nonsensical combinations can work):

When making a negative prompt, you can check if you're on the right track by pasting it by itself in the positive prompt.

The above is from 1.5 with the following in the positive:

quote:

lowres, low resolution, thumbnail, preview, template, deformed, distorted, beginner, amateur, low quality, bad quality, worst quality, overexposed, underexposed, aberration, video game, screenshot, captcha, forum, watermark, meme, text, signature, blurry, cropped, poorly drawn, glitch, aliasing, censored

# ? Aug 2, 2023 19:32

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

RIP Syndrome posted:

It usually helps to specify the style you want, e.g. "photo". Otherwise you get whatever. You can be more specific about the imaging tech, for instance try "35mm", "kodachrome", camera models and so on. E.g. "hdr polaroid" (nonsensical combinations can work):

When making a negative prompt, you can check if you're on the right track by pasting it by itself in the positive prompt.

The above is from 1.5 with the following in the positive:

I haven't had problems with 1.5, 2.1 or Bing, this is specifically an SDXL issue or whatever the hell that workflow on comfyUI is. Like it generally leans really hard into realistic that you don't need to add more realism words. Maybe SDXL is just more balanced?

# ? Aug 2, 2023 19:39

Hadlock: Nov 9, 2004

pixaal posted:

I'm playing with SDXL 1.0 with the 1.0 VAE because that's what I was uploading earlier and then had to troubleshoot some errors

It really keeps trying to go back to this comicbook look, not that I don't like it but I don't always want it. maybe someone has an idea

I'm sure it's in the mess of settings still at default

What does ComfyUI's "settings" or "config" export format layout look like. Would be neat if they had a QR code export option for that. These workflows look like they can/will get complex fast

Maybe in your negative prompts add "jack kirby" and "Sam Kieth" and "Todd McFarlane" who are all comic book artists; also "marvel comics" and "dc comics"

# ? Aug 2, 2023 19:46

RIP Syndrome: Feb 24, 2016

Two other things that give me bad results (often naive sketch/comic styles) with SDXL and a1111:

- Generating at less than 1024x1024.
- Trying to upscale with the refiner. I've only gotten it to work well at 1:1.

But I don't know about ComfyUI. Maybe there's some implicit resampling step going wrong. I'd leave the refiner out altogether while looking for the cause, as it's not really critical.

# ? Aug 2, 2023 19:50

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Hadlock posted:

What does ComfyUI's "settings" or "config" export format layout look like. Would be neat if they had a QR code export option for that. These workflows look like they can/will get complex fast

Maybe in your negative prompts add "jack kirby" and "Sam Kieth" and "Todd McFarlane" who are all comic book artists; also "marvel comics" and "dc comics"

it's a Json that workflow is the 1.0 from here https://github.com/SytanSD/Sytan-SDXL-ComfyUI It's not even the most complicated one I looked at, but that one needed an addon I couldn't get installed in colo and will have to take a look at on hardware tonight. This guy has a ton of weird ones
this is their example not mine:
https://github.com/wyrde/wyrde-comfyui-workflows/tree/main

e: What the one I'm using is doing is 20 steps with the base one then passing that off to the refiner for the final 10 steps then it makes the image on the left then it passes that off to an upscaler and makes the image on the right
e2: getting kind of busy with actual stuff I need to do but photo does seem to be what was missing. XL seems much more neutral than previous models. I've just been able to apply so much to both 1.5 and 2.1 I didn't think XL would be this different (or maybe it's an artifact of how comfy is working)

pixaal fucked around with this message at 20:04 on Aug 2, 2023

# ? Aug 2, 2023 19:51

Killer robot: Sep 6, 2010; I was having the most wonderful dream. I think you were in it!; Pillbug

Silly simple prompts in DucHaiten-Journey.

satan's own chihuahua, yapping at the gates of hell

satan's best buddy, waiting at the gates of hell for him to get off work:

I get a 1990s adventure game vibe from the second, and 2010s HD remake from the third.

# ? Aug 2, 2023 21:12

Sedgr: Sep 16, 2007; Neat!

Sedgr fucked around with this message at 21:40 on Aug 2, 2023

# ? Aug 2, 2023 21:17

KakerMix: Apr 8, 2004; 8.2 M.P.G.

Weeks ago when Meta's music AI was being talked about someone mentioned a sound effects maker would be cool. Well:

https://ai.meta.com/blog/audiocraft-musicgen-audiogen-encodec-generative-ai-audio/

The picture itself isn't too wildly special, but it is the result of me loving around with the [url=https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin#how-to-install]photoshop plugin[./url] using Stable Diffusion XL to figure out how to use it. Lots better than the previous one with Krita. Not only because it's Photoshop and I'm far more familiar with it, but also in how it handles using SD inside an image editing suite.

KakerMix fucked around with this message at 00:17 on Aug 3, 2023

# ? Aug 3, 2023 00:12