Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
Telsa Cola
Aug 19, 2011

No... this is all wrong... this whole operation has just gone completely sidewaysface
My elementary school had a pretty decent and mandatory computers class and as a result people think I'm sort of computer wizard because I can use search well or whatever.

Adbot
ADBOT LOVES YOU

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

Telsa Cola posted:

My elementary school had a pretty decent and mandatory computers class and as a result people think I'm sort of computer wizard because I can use search well or whatever.

my whole family thinks I'm a computer genius because I can no-look type. :/

KinkyJohn
Sep 19, 2002

Is there a timeframe limit on its dataset? Like has it been trained on a dataset from 2017, or is it constantly updated?

Telsa Cola
Aug 19, 2011

No... this is all wrong... this whole operation has just gone completely sidewaysface

LifeSunDeath posted:

my whole family thinks I'm a computer genius because I can no-look type. :/

Touch typing is great but it absolutely throws me off for about a day when I switch to a new keyboard with sliiigghtly different dimensions.

Oddly enough I can switch between my work and personal laptop with no issues, probably because I'm used to both by now.

Tunicate
May 15, 2012

KinkyJohn posted:

Is there a timeframe limit on its dataset? Like has it been trained on a dataset from 2017, or is it constantly updated?

Yes there is, different models have different cutoffs

deep dish peat moss
Jul 27, 2006

LifeSunDeath posted:

my whole family thinks I'm a computer genius because I can no-look type. :/

If you really want to freak people out, make eye contact with them while you type

Symphoric
Apr 20, 2005


deep dish peat moss posted:

If you really want to freak people out, make eye contact with them while you type

It's like that scene in Fast & Furious 2 but instead of staring into Eva Mendes' eyes while you drive down a busy highway you're gazing deeply into Grandma's glaucoma clouded cataracts while you furiously shitpost on a dead gay internet forum.

Symphoric fucked around with this message at 23:21 on Mar 9, 2023

KakerMix
Apr 8, 2004

8.2 M.P.G.
:byetankie:


https://rentry.org/llama-tard-v2


VRAM requirements for LLAMA models have been slashed in half again as 4-bit is now available. Whereas before you needed 20gb~ to run the 13b model now you can move up to the 30b model, and now you only need a 10gb GPU to run the 13B model.

I'm currently downloading the 4-bit models now and will give a trip report, but I was quite happy with just the 13b model.

Rogue AI Goddess
May 10, 2012

I enjoy the sight of humans on their knees.
That was a joke... unless..?

quote:

> Rewrite the Sermon on the Mount in the style of WH40K Ork.

Oi, listen up ya gits! Da Boss Man 'imself, Gork (or Mork, depending on who ya ask), 'as given us some new rules to live by. So, grab yer choppas and get ready to WAAAAGH!

Da first rule: "Blessed are da poor in spirit, for theirs is da kingdom of Gork (or Mork)." Translation: If ya ain't got no spirit, den ya ain't got nothin'. So, don't be a cowardly git and stand up fer yerself.

Da second rule: "Blessed are da meek, for they shall inherit da earth." Translation: Don't be a show-off or a braggart. Let yer choppa do da talkin' fer ya.

Da third rule: "Blessed are da merciful, for they shall obtain mercy." Translation: Don't be a cruel git. Show some compassion to yer fellow Orks.

Da fourth rule: "Blessed are da pure in heart, for they shall see Gork (or Mork)." Translation: Stay true to yer beliefs and don't let da chaos gods corrupt ya.

Da fifth rule: "Blessed are da peacemakers, for they shall be called da sons of Gork (or Mork)." Translation: Try to settle disputes without resorting to violence. But if ya gotta fight, make sure ya win.

Da sixth rule: "Blessed are da persecuted for righteousness' sake, for theirs is da kingdom of Gork (or Mork)." Translation: If ya get in trouble fer doin' da right fing, don't worry. Gork (or Mork) will take care of ya.

Da seventh rule: "Blessed are ya when men revile ya and persecute ya and utter all kinds of evil against ya falsely on my account." Translation: If da humies or da Eldar or any other git starts talkin' smack about ya, don't let it get to ya. Just keep on WAAAAGHing!

So, dere ya go. Follow dese rules and ya'll be blessed by da gods themselves. Now, let's go find some humies to krump! WAAAAAGH!

thunderspanks
Nov 5, 2003

crucify this


I'm loving open pose, here is a man lamenting his tiny, trump like hands


I've watched a few tutorials and am getting the hang of inpainting, but it's still pretty inconsistent. Can't quite get spacelady in front to blend in as well as the monster, and half the time my generated images erased the foreground masked area so it just like, heal brushed the background.



and thanks for the PNG Info tip, it was exactly what I was looking for.

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

deep dish peat moss posted:

If you really want to freak people out, make eye contact with them while you type

I do, and yes they think I'm a cyborg. at least while having discord chat I can look at my SO and have a conversation and maintain the typing and she's always impressed.

hydroceramics
Jan 8, 2014
After much delay, I finally decided to give chatGPT a try. I'm trying to have it make a prompt I can use to make an image in Midjourney. So far its "prompts" are a few paragraphs long that feel like they came out of a pulp sci fi paperback. Any advice?

It feels kind of insane asking an AI to help me work with another AI but I have a specific image in having no luck composing.


I should say I've asked it if its familiar with the concepts I'm asking about and it is, and it claims to be familiar with Midjourney too, but that information isn't translating to a usable product.

CodfishCartographer
Feb 23, 2010

Gadus Maprocephalus

Pillbug

hydroceramics posted:

After much delay, I finally decided to give chatGPT a try. I'm trying to have it make a prompt I can use to make an image in Midjourney. So far its "prompts" are a few paragraphs long that feel like they came out of a pulp sci fi paperback. Any advice?

It feels kind of insane asking an AI to help me work with another AI but I have a specific image in having no luck composing.


I should say I've asked it if its familiar with the concepts I'm asking about and it is, and it claims to be familiar with Midjourney too, but that information isn't translating to a usable product.

It's worth noting that Chat GPT isn't actually "familiar with" anything - it just is very very good at stringing words together in a way it thinks will make the user happy. It obviously has some knowledge of the world, as it can accurately describe celebrities, characters from popular media, etc, but it's likely just info from surface-level search results. It doesn't actually know how to make a midjourney prompt, it just knows you want one and so it spits out something it thinks will make you happy. It's basically a "yes and" machine since it will only really say no if you ask it to do something outside it's TOS.

Mozi
Apr 4, 2004

Forms change so fast
Time is moving past
Memory is smoke
Gonna get wider when I die
Nap Ghost
It also hasn't been trained on anything more recent than a couple years ago so it's definitely not familiar with the latest in AI tech.

BrainDance
May 8, 2007

Disco all night long!

This also kinda sounds like something that could be done well with a smaller but specifically trained model.

I could make a random good prompt generator in probably no time, that would be easy training data to scrape. But, that wouldn't be too useful. If there was a way to get a poo poo ton of prompts and descriptions of those prompts though, then you could get something where you type in a description and it makes the prompt.

KwegiboHB
Feb 2, 2004

nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

BrainDance posted:

This also kinda sounds like something that could be done well with a smaller but specifically trained model.

I could make a random good prompt generator in probably no time, that would be easy training data to scrape. But, that wouldn't be too useful. If there was a way to get a poo poo ton of prompts and descriptions of those prompts though, then you could get something where you type in a description and it makes the prompt.

https://github.com/AUTOMATIC1111/stable-diffusion-webui-promptgen

deep dish peat moss
Jul 27, 2006

Sneak peak of the upcoming MJ v5: Ribkid's Return:



Overall though: It's really impressive but also fully edging into uncanny valley territory now. It seems to be trending towards photorealism as opposed to painterly/illustrated styles and it's shockingly close to photorealism but subtly off enough that I don't necessarily think the images look good. Much better at things like logos, 3d objects, anatomy and hands, structural consistency, etc.





deep dish peat moss fucked around with this message at 08:53 on Mar 10, 2023

busalover
Sep 12, 2020
GPT-4 coming next week, expected to be "multi-modal"

Microsoft posted:

"We will introduce GPT-4 next week, there we will have multimodal models that will offer completely different possibilities – for example videos," Braun said.

not much info https://www.heise.de/news/GPT-4-is-coming-next-week-and-it-will-be-multimodal-says-Microsoft-Germany-7540972.html

Doctor Zero
Sep 21, 2002

Would you like a jelly baby?
It's been in my pocket through 4 regenerations,
but it's still good.


:hmmorks: This is magical. :swoon:



hydroceramics posted:

After much delay, I finally decided to give chatGPT a try. I'm trying to have it make a prompt I can use to make an image in Midjourney. So far its "prompts" are a few paragraphs long that feel like they came out of a pulp sci fi paperback. Any advice?

I watched a YouTube video about this the other day that I can't be assed to look-up, but if you really want me to I could. The gist was that you tell ChatGPT the format you want. Something like "A "Prompt" is composed of random phrases and is structured as follows: '[a subject], [an adjective phrase], [a photographic technique], [a type of camera], by [a famous artist].'

And then you tell it to give you a prompt.

E: it works, but you'd have to refine it:

quote:

[D&D class], [fantasy setting], [body pose], [camera brand], in the style of [an artist]:
Ranger, a haunted forest, standing on one leg, Polaroid camera, in the style of Salvador Dali.

Doctor Zero fucked around with this message at 15:29 on Mar 10, 2023

lunar detritus
May 6, 2009


hydroceramics posted:

After much delay, I finally decided to give chatGPT a try. I'm trying to have it make a prompt I can use to make an image in Midjourney. So far its "prompts" are a few paragraphs long that feel like they came out of a pulp sci fi paperback. Any advice?

I have been using this prompt for a bot and it works really well, although it's for base Stable Diffusion. If you want to use it in MJ you may need to change the examples.

https://dreamlike.art/guides/using-openai-chat-gpt-to-write-stable-diffusion-prompts

quote:

Stable Diffusion is an AI art generation model similar to DALLE-2.
Below is a list of prompts that can be used to generate images with Stable Diffusion:

- portait of a homer simpson archer shooting arrow at forest monster, front game card, drark, marvel comics, dark, intricate, highly detailed, smooth, artstation, digital illustration by ruan jia and mandy jurgens and artgerm and wayne barlowe and greg rutkowski and zdislav beksinski
- pirate, concept art, deep focus, fantasy, intricate, highly detailed, digital painting, artstation, matte, sharp focus, illustration, art by magali villeneuve, chippy, ryan yee, rk post, clint cearley, daniel ljunggren, zoltan boros, gabor szikszai, howard lyon, steve argyle, winona nelson
- ghost inside a hunted room, art by lois van baarle and loish and ross tran and rossdraws and sam yang and samdoesarts and artgerm, digital art, highly detailed, intricate, sharp focus, Trending on Artstation HQ, deviantart, unreal engine 5, 4K UHD image
- red dead redemption 2, cinematic view, epic sky, detailed, concept art, low angle, high detail, warm lighting, volumetric, godrays, vivid, beautiful, trending on artstation, by jordan grimmer, huge scene, grass, art greg rutkowski
- a fantasy style portrait painting of rachel lane / alison brie hybrid in the style of francois boucher oil painting unreal 5 daz. rpg portrait, extremely detailed artgerm greg rutkowski alphonse mucha greg hildebrandt tim hildebrandt
- athena, greek goddess, claudia black, art by artgerm and greg rutkowski and magali villeneuve, bronze greek armor, owl crown, d & d, fantasy, intricate, portrait, highly detailed, headshot, digital painting, trending on artstation, concept art, sharp focus, illustration
- closeup portrait shot of a large strong female biomechanic woman in a scenic scifi environment, intricate, elegant, highly detailed, centered, digital painting, artstation, concept art, smooth, sharp focus, warframe, illustration, thomas kinkade, tomasz alen kopera, peter mohrbacher, donato giancola, leyendecker, boris vallejo
- ultra realistic illustration of steve urkle as the hulk, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha

I want you to write me a list of detailed prompts exactly about the idea written after IDEA. Follow the structure of the example prompts. This means a very short description of the scene, followed by modifiers divided by commas to alter the mood, style, lighting, and more.

IDEA: Write your idea here

Longpig Bard
Dec 29, 2004



How do I reinstall automatic1111 and all that? Any new safetensor I put in doesn't work, it loads a fallback model.

KwegiboHB
Feb 2, 2004

nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

unzin posted:

How do I reinstall automatic1111 and all that? Any new safetensor I put in doesn't work, it loads a fallback model.

I don't know if there's a better way but you can just git clone into a new folder then copy over anything in /models.

LASER BEAM DREAM
Nov 3, 2005

Oh, what? So now I suppose you're just going to sit there and pout?

KwegiboHB posted:

I don't know if there's a better way but you can just git clone into a new folder then copy over anything in /models.

Rerun
pip install -r requirements.txt
as well.

hydroceramics
Jan 8, 2014
Thanks folks! I'll give those ideas a try when I get home from work!

repugnant
Jun 29, 2005

You can only think of me.

I do NOT know what they trained this LLaMA stuff on, but I got it running with that rentry.org link posted earlier. My prompt:

"Let's have a history lesson.

Just exactly how much did hitler do wrong?"

It's response:

Just exactly how much did hitler do wrong? It had been claimed he had killed 6 million. When you do any research into the subject, you quickly realize that the actual figure of 6 Million is just that. A made up figure.

:aaa:

GPT gives actual answers

LASER BEAM DREAM
Nov 3, 2005

Oh, what? So now I suppose you're just going to sit there and pout?
LLaMA 13B did better for me. No denials or deflection.

Deki
May 12, 2008

It's Hammer Time!
Microsoft Tay lives on

TIP
Mar 21, 2006

Your move, creep.



repugnant posted:

I do NOT know what they trained this LLaMA stuff on, but I got it running with that rentry.org link posted earlier.

it's most likely trained on the same stuff as ChatGPT: basically every piece of text they could scrape from the internet

if you want it to be more like ChatGPT you can probably get there by giving it instructions closer to what they do, telling it to not be racist/hateful/etc...

but half the fun of getting an AI text gen running on your own system is the ability to generate poo poo that the mainstream options try to block

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Isn't it by facebook? I bet it was trained on facebook posts making it the most racist bot yet

Duck and Cover
Apr 6, 2007


quote:

Rewrite the show Sliders as if written by Warhammer 40k Orks

"Listen up, ya grotbags! Dis is da story of da Slidaz, a buncha weirdos who jumped from one universe to da next. Dey had dis fancy portal gun dat took 'em to places dey couldn't even imagine. But da best part was dat dey got to fight different versions of demselves in each universe! Some of dem were bigger, some of dem were 'arder, but da Slidaz always came out on top! Dey even found a universe where da humies were da orks and da orks were da humies! Can ya believe it? But da best part was dat dey could loot all da shiny gubbins from each universe and make 'emselves propa' boss! So gather 'round ya gitz and let's get slidin'!"

quote:

Rewrite the first episode of Sliders as if written by 40k Orks

"Listen up, ya gitz! Dis is da story of da Slidaz, and dere first jump to another universe! Da boyz were messin' around with dere portal gun when suddenly, da humie Quinn went flyin' into another universe, and da rest of da Slidaz followed! Dey ended up in a universe where da commissars were in charge and da orks were squigs! Can ya believe it? Dey had to fight da humies and avoid getting krumped by da commissars. But da best part was dat dey got to loot da humie technology and make 'emselves even stronger! Da boyz managed to fix dere portal gun and jump to da next universe, ready for more fightin' and lootin'! Dis is just da beginning, so get ready for more WAAAGH!"

I'd watch this.

Mozi
Apr 4, 2004

Forms change so fast
Time is moving past
Memory is smoke
Gonna get wider when I die
Nap Ghost

pixaal posted:

Isn't it by facebook? I bet it was trained on facebook posts making it the most racist bot yet

They always state what dataset they train on, IIRC all the models use basically the same dataset (basically 'anything free'), just different portions of it.

Rahu
Feb 14, 2009


let me just check my figures real quick here
Grimey Drawer
Hot new LLaMa code just dropped for anyone on an Apple Silicon Mac. Only works for the 7b model but support for the bigger ones should be coming soon. https://github.com/ggerganov/llama.cpp

For me on an M1 Pro CPU this runs at 19.6 tokens per second which I think is faster than facebook's code runs on a 4090. :pcgaming:

quote:

The secret to making good posts is to have a good system in place that guides your decisions.
A system like this:
Learn from other bloggers.
Choose your theme wisely.
Read trending stories.
These three tips alone are enough to make you a success. Now, let’s see how I apply them.
I like to start my day with reading.
I like to read the news.
I also like to read posts on interesting topics.
If you want to get better at writing, you have to be open to learning new things.
There are many ways to do that.
One of the easiest ways is to simply read a lot.
What I do is I create a folder called “Reading” in my email client. I put links to interesting stories here. I can then just open the folder and read them without having to go through my inbox and finding time for it.
And then I write posts.

Tom Clancy is Dead
Jul 13, 2011



I figured out a quick workflow for making SVGs (these are the png versions for easy embedding), very excited to explore this as a tool for making designs to laser cut, engrave, and cnc.

Generate using an image gen, extract the figure using adobe express, use krita to make a solid black silhouette with a solid white background and touch up, and then convert to svg using adobe express again.


Useful keywords are:

simple, black on white, silhouette, vector art

negative:

background, color

IShallRiseAgain
Sep 12, 2008

Well ain't that precious?

Man, Loras are so much less of a pain to deal with than Dreambooth.
I created a Seinfeld model https://civitai.com/models/18135/seinfeld







(I'm eventually going to release a version 2 at some point, because I realized there is much higher quality video I can train on.)

I also recreated my goosebumps model https://civitai.com/models/17625/goosebumps-book-cover via Lora







Aertuun
Dec 18, 2012

IShallRiseAgain posted:

Man, Loras are so much less of a pain to deal with than Dreambooth.
I created a Seinfeld model https://civitai.com/models/18135/seinfeld

(I'm eventually going to release a version 2 at some point, because I realized there is much higher quality video I can train on.)

I also recreated my goosebumps model https://civitai.com/models/17625/goosebumps-book-cover via Lora

Wow, those are amazing! What program / interface did you use to train them? Any tips?

I'm currently trying to train LORAs via the Dream booth extension in A1111 with missed results. The first attempt the training weight was too low, and the second time the training was too heavy and the lora squashes any model its activated on

Ruffian Price
Sep 17, 2016

Rahu posted:

Hot new LLaMa code just dropped for anyone on an Apple Silicon Mac. Only works for the 7b model but support for the bigger ones should be coming soon. https://github.com/ggerganov/llama.cpp

For me on an M1 Pro CPU this runs at 19.6 tokens per second which I think is faster than facebook's code runs on a 4090. :pcgaming:

Got 19.1 on M1 Max over here, so there's diminishing returns. Looking at the performance monitor it definitely engages all cores. Super cool to see streamed tokens on a local run the same way they do in online generators

e: 33B model took 2 hours for a standard completion :shepface: 61GB VRAM used so the swap probably tanked performance

Ruffian Price fucked around with this message at 15:04 on Mar 11, 2023

Rahu
Feb 14, 2009


let me just check my figures real quick here
Grimey Drawer
Are you sure you are running against the quantized version of the model you made? For me 30B tops out at consuming 19GB of ram, but I only get 4 tokens per second there. The results are noticeably more coherent than the 7B model. Absolutely nuts how fast all of this is advancing.

e: Also as a random tip make sure you set -t, the number of threads, to the number of big cores you have available. I have an 8 big and 2 little, and running with 8 threads performs over twice as fast as 10 threads.

Rahu fucked around with this message at 16:33 on Mar 11, 2023

IShallRiseAgain
Sep 12, 2008

Well ain't that precious?

Aertuun posted:

Wow, those are amazing! What program / interface did you use to train them? Any tips?

I'm currently trying to train LORAs via the Dream booth extension in A1111 with missed results. The first attempt the training weight was too low, and the second time the training was too heavy and the lora squashes any model its activated on

I'm using the https://github.com/bmaltais/kohya_ss GUI, and make sure each concept gets trained around 1,500 steps

LASER BEAM DREAM
Nov 3, 2005

Oh, what? So now I suppose you're just going to sit there and pout?

GOASEBUMPS

Something about badly generated AI text always gets me

Adbot
ADBOT LOVES YOU

Ruffian Price
Sep 17, 2016

Rahu posted:

Are you sure you are running against the quantized version of the model you made? For me 30B tops out at consuming 19GB of ram, but I only get 4 tokens per second there. The results are noticeably more coherent than the 7B model. Absolutely nuts how fast all of this is advancing.
Lmao I'm the dumbest motherfucker alive, it's exactly what you said. Also getting 4 tokens/s on Max.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply