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

HORSE-SLAUGHTERER: Nov 11, 2020; H O R S E - S L A U G H T E R E R

Hi I'm late to the party with this stuff, spent the last couple of days messing with bing image creator and ended up obsessing over gothic/steampunk/cyberpunk angels (because if nothing else AI is good at generating tropey generic stuff).

The prompts were some variation of �neon winged goddess flying over a gothic smoky steampunk city� - some of the best ones had the word "fractal" in the string also.

# ? Oct 18, 2023 20:52

Adbot: ADBOT LOVES YOU

# ? May 29, 2024 16:05

Tarkus: Aug 27, 2000

Those are really nice!

# ? Oct 18, 2023 20:53

Sedgr: Sep 16, 2007; Neat!

# ? Oct 18, 2023 21:02

Mordiceius: Nov 10, 2007; If you think calling me names is gonna get a rise out me, think again. I like my life as an idiot!

I feel like the Bing image creator had a gently caress ton of holes plugged last night because so many of my old prompts are just endless eggdog.

# ? Oct 18, 2023 21:08

have you seen my baby: Nov 22, 2009

# ? Oct 18, 2023 22:31

XYZAB: Jun 29, 2003; HNNNNNGG!!

Can I get a hat wobble? :smug:

# ? Oct 18, 2023 22:44

TIP: Mar 21, 2006; Your move, creep.

# ? Oct 18, 2023 22:51

XYZAB: Jun 29, 2003; HNNNNNGG!!

hELL YEAH

# ? Oct 18, 2023 23:24

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

# ? Oct 18, 2023 23:26

Montague Tigg: Mar 23, 2008; Previously, on "Ronnie Likes Data":

KinkyJohn posted:

yogi bear movie poster with booboo standing in front of yogi bear, the title says "Great things come in bears"

Gets me dogged in bing

# ? Oct 19, 2023 00:29

TIP: Mar 21, 2006; Your move, creep.

Montague Tigg posted:

the joke was that they're describing the actual poster for yogi bear

and the real thing is so much worse than any of your generations :lmao:

# ? Oct 19, 2023 00:31

Montague Tigg: Mar 23, 2008; Previously, on "Ronnie Likes Data":

# ? Oct 19, 2023 00:32

Fuckstick: Nov 30, 2000

Seen this one in a Facebook group and it made me crack up. This prompt doesn�t disappoint. First photo is from the group, the second one is what I got when I used the same prompt.

prompt: "obese Walmart people shopping carts, seagulls, chaos, fire department"

# ? Oct 19, 2023 01:10

Tarkus: Aug 27, 2000

I find it interesting that it doesn't generate a proper manufacturing machine when asked. I would wager that the issue is that most machines don't have a pre-defined configuration.

If I ask specifically for a lathe, it seems to get it closer than others

It does not however know what machines in a machine shop look like.

# ? Oct 19, 2023 02:59

Soulhunter: Dec 2, 2005

I heard there's some new Five Guys at Freddy's movie coming soon

# ? Oct 19, 2023 03:56

TIP: Mar 21, 2006; Your move, creep.

where's the hamburgers? :confused:

# ? Oct 19, 2023 04:02

Soulhunter: Dec 2, 2005

oh no somebody must've dropped the burgers, better move the couch and bend over to look for them

Soulhunter fucked around with this message at 04:17 on Oct 19, 2023

# ? Oct 19, 2023 04:07

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

# ? Oct 19, 2023 04:36

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

If someone can figure out how to get bing/DALLE3 to generate 'a smiling happy family readily eating rat nuggets in a McDonalds franchise owned by H.R. Giger' please help my family is starving

# ? Oct 19, 2023 04:56

TIP: Mar 21, 2006; Your move, creep.

Mercury_Storm posted:

If someone can figure out how to get bing/DALLE3 to generate 'a smiling happy family readily eating rat nuggets in a McDonalds franchise owned by H.R. Giger' please help my family is starving

a smiling happy family readily eating rat nuggets in a fast food restaurant with golden arches in a franchise owned by H.R. Giger

# ? Oct 19, 2023 05:21

SniperWoreConverse: Mar 20, 2010; Gun Saliva

these are strong terms that seem to try and override each other

dehumanize yourself kill you're famil

# ? Oct 19, 2023 05:25

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

dear god lol, thank you

# ? Oct 19, 2023 05:27

SniperWoreConverse: Mar 20, 2010; Gun Saliva

Will u worship at the altar of the lord your mcdonad

# ? Oct 19, 2023 05:35

Montague Tigg: Mar 23, 2008; Previously, on "Ronnie Likes Data":

the god of fast food

# ? Oct 19, 2023 05:50

SniperWoreConverse: Mar 20, 2010; Gun Saliva

# ? Oct 19, 2023 06:01

KinkyJohn: Sep 19, 2002

Turns out Dalle-e-3 doesn't want me to generate these prompts, so I'm just going to quote myself from the day that MJ v5 was released:

KinkyJohn posted:

Retrying some of my old MJ v3 prompts

Mcdonalds Megachurch architectural renderings:

Ronald McJesus handing out flyers in the parking lot:

# ? Oct 19, 2023 08:00

TIP: Mar 21, 2006; Your move, creep.

huh, Bing had no issue pumping out a bunch of images based on those prompts for me

# ? Oct 19, 2023 08:18

PaleFigure: Sep 3, 2007; the other white meat

I am legitimately baffled by what causes Bing to dog an image or not. It clearly isn't just the words used in the prompt, because re-submitting the same prompt can produce a combo of generated and dogged results. I understand they have to have *some* protections in place since this is a free, publicly accessible tool, but I wish it wasn't quite so obtuse, or seemingly just plain *random*.

It's especially frustrating when it seems to break it's own rules and produce something borderline violent or adult from a perfectly normal prompt. Case in point; I gave it the prompt "A pair of satyrs dance amongst falling autumn leaves", and on multiple iterations it gave me (admittedly tasteful) goat-girl boobs.

:nws:

obviously

# ? Oct 19, 2023 08:34

Small Strange Bird: Sep 22, 2006; Merci, chaton!

I think you get the eggdog if it produces an image that's been generated using an element that MS has decided is tagged as 'unsafe' (even if there was nothing in your prompt that you'd think would be), and it catches it before displaying - if all four in a run are caught, you get egged. I was using BIC on my phone yesterday when it was running slow, and for a split-second four images would flash up before being culled down to less. I've often found that you can send an eggdogged prompt again and it'll deliver something, so maybe it worked from a similar but non-tagged element.

If you have something in your prompt that's prohibited it gets pissy almost immediately, so it probably doesn't even pass it on from Bing to Dall-E.

It's all potluck, though. I've had images giving me unrequested bare butts and breasts, nip-slips, blood and gore...

# ? Oct 19, 2023 10:10

ymgve: Jan 2, 2004; Offensive Clock

I think it might be mostly copyright issues - I never had problems before I tried making referential animal movie posters. I guess 90% of the time it is �unsafe for us at Bing and our copyright lawyers�

# ? Oct 19, 2023 13:55

Mr Luxury Yacht: Apr 16, 2012

So Nvidia released an official extension for Stable Diffusion/Automatic 111 for Tensor RT.

There's a couple things with it that are a bit of a pain. For one, you need to compile a separate engine for each checkpoint specifying the min/max height, width, and batch for the images. If you want to use hiresfix then the width/height has to be large enough to include the max size you upscale images to, which can make the engine model take a while to generate and get a bit big. If you use hiresfix, there's also a noticeable delay for each image as it separately loads the engine to upscale. LORAs also have to be baked into the engine at generation.

However if you're the kind of person that generates a ton of images and then individually upscales the best ones don't sleep on it. It literally doubled my iterations per second. For 512x512 on a 4080 I'm getting well over 40 iterations per second now with SD 1.5 checkpoints. Haven't tried SDXL yet but supposedly it gives similar levels of performance increase.

# ? Oct 19, 2023 14:21

Unkempt: May 24, 2003; ...perfect spiral, scientists are still figuring it out...

Nyan Bread posted:

If you're looking to dive into the weird and wonderful world of AI art generation using Stable Diffusion on your local PC and have a video card with at least 8GB VRAM, this weekend might be the best time to start with the newly released higher-res SDXL dataset. Here's a bunch of tips to help you on your way to setting up a basic workflow in the very flexible node-based browser-based editor ComfyUI:

Download and install Stability Matrix AI suite manager that will be used for downloading custom SD models and further customizations, as well as updating ComfyUI - https://github.com/LykosAI/StabilityMatrix/releases

Use the Packages tab to download the latest ComfyUI. Once finished it will be integrated into the Stability Matrix suite to allow using the Model Browser tab to download the trained datasets directly from https://civitai.com

Now onto the actual ComfyUI editor where the AI art generation will take place. First get the ComfyUI-Manager plugin:
https://github.com/ltdrdata/ComfyUI-Manager/archive/refs/heads/main.zip

Extract, and copy the ComfyUI-Manager-main folder into your "\StabilityMatrix\Data\Packages\ComfyUI\custom_nodes" directory. Relaunch ComfyUI and it'll download the dependencies and the Manager new button will show up in your editor.

Within the Manager click Install Custom Nodes and use the search field to install the following plugins for painless use of this noodle-based editor: (also make sure to relaunch the app so it gets the dependencies)
ComfyUI Impact Pack
Efficienty Nodes for ComfyUI
ComfyUI-Bmad-DitryUndoRedo

From the Model Browser tab of Stability Matrix I recommend you download the following startup components:
SD XL (get both the VAE Fix and the refiner VAE Fix)
SDXL VAE
Detail Tweaker XL
realESRGAN_x4+HAT (then copy realesrganX4HAT_v10.ckpt from your \StabilityMatrix\Data\Models\StableDiffusion\ directory into the \StabilityMatrix\Data\Models\ESRGAN directory)

Once those are downloaded you should be good to go as far as having all the raw image data to manipulate and the tools to do it with offline. Now here's some basic components of the actual ComfyUI workflow that you can build upon:

The compact "all-in-one" loader is used for loading the base model, the refiner, specifying the size of the generation as well as connecting to the Low-Rank Adaptation of Large Language Model (LoRA) stack which are like contextual modifiers or "smart filters" for the generated image:

The sampler is the engine of AI generation that performs functions on the dataset to come up with a coherent image. You can play with these settings to generate wildly different images from the same promt:

FaceDetailer is an automatic inpainting module that can fix messed up faces/eyes/hands but also slightly alters the entire image quality since it does contextual blending. Hopefully a smarter one comes out soon that only blends the image around the area of the mask:

Upscaler increases the size of the image heuristically and is pretty fast, more upscalers can be downloaded from https://huggingface.co

Lastly, you can import and export workflows in ComfyUI as .json files - here's the one in the example, enjoy your journey! https://pastebin.com/raw/EHvnh8y7

Trying to follow this lovely post and it's good until 'From the Model Browser tab of Stability Matrix I recommend you download the following startup components'. I'm just getting 'Civitai can't be reached right now' and have been since yesterday, any ideas?

# ? Oct 19, 2023 15:01

null_pointer: Nov 9, 2004; Center in, pull back. Stop. Track 45 right. Stop. Center and stop.

First time trying Bing, asked it to create a sports team logo for my very much non-sporting corporate scrum team that handles "accounts". Let's just say that Bing does not have a problem dropping the "o" from the word.

:eyepop:

# ? Oct 19, 2023 15:35

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

Text constantly gets things dropped and it really has very little idea what it writes there. I don't think it even attempts to read text output to sterilize!

I was making a chicken cooking and got a bottle of cock(acr Koun)