Peter Baylies @pbaylies Twitter profile

Pinned Tweet

Peter Baylies

4 years

Ok folks, I now have an updated repo and new WikiArt models! New repo here: Models here: Thanks to @dvsmethid @cschan9 @davidstap @robertluxemburg @Veqtor @theshawwn for their repos and their work in this space!

Peter Baylies

@pbaylies

4 years

@MichaelFriese10 @ak92501 @theshawwn Ok folks, *now* I have a WikiArt conditional model released and a repo up! Model here: Repo here: Thanks to @davidstap @robertluxemburg @Veqtor @theshawwn for their repos and their work in this space!

9

24

117

10

49

247

Last Seen Profiles

@GeenyMariee

@AlbinaKealy914

@fox112345

@moncheriyoo

@haleypauthor

@qi_atr

@myjacksdck

@witchvitt

@AlonePodcast

@khiawings

@mariepoulin

@travisbeacham

@Jeden_z_wielu

@javillanoabril

@kooo_gi

@ed_reys

@docencik1234

@anc_ribeiro

@Arion7_

@MORANA_DOMINA

@SydWnyn

@l3nnly

@den_run_ai

@Brilhante_rsb

@ArianeGroup

@d_pkwd

@BellaVanvo28002

@DnQx91

@betr

@radh_al52040

@jesskellynt

@alaslamnjm16

@aydaleium

@innafree_69

@EUatUN

@mxboyshockey

Peter Baylies

@pbaylies

2 years

This is a lie. As the guy who wrote a good chunk of the code that these fake images were likely generated with, I do not approve this use or this bad take. The way the old StyleGAN encoder worked was to intentionally try to generate an image as close to a real image as possible.

Bob Flynn - same on bloosky

@bobjinx

2 years

This is what artists are referring to when they say the 'machine-learning' of text-to-sludge leverages stolen art. source (and a great critique of tech-hype)

47

916

10K

14

128

1K

Peter Baylies

@pbaylies

5 years

Watch the Mona Lisa turn her head, in today's edition of stupid StyleGAN encoder tricks... #StyleGAN cc: @aarlo @Zergfriend @neurokinetikz @calamardh @roadrunning01 @Miles_Brundage

17

296

928

Peter Baylies

@pbaylies

1 month

@MysteryGuitarM If you assume that some sort of Moore's Law scaling holds long enough ("doubles every two years"), then 2 * (log(570000000) / log(2)) = upload yourself to your Macbook equivalent in less than 60 years?

18

12

739

Peter Baylies

@pbaylies

3 years

@andrew_n_carr Ok but don't forget about the derivative of ln(n^2) 2/n

2

9

306

Peter Baylies

@pbaylies

2 years

That is not the same thing as the website this person does not exist at all which is why this is misleading incorrect and outright false information and really you should consider deleting it and your account on Twitter.

3

6

288

Peter Baylies

@pbaylies

2 years

@diracdelta @matthew_d_green Little Qa'taq Tables, we'd call him...

3

8

181

Peter Baylies

@pbaylies

2 years

@nixcraft @DevDminGod Next step, forcing it to debug and fix its own code...

6

4

141

Peter Baylies

@pbaylies

2 years

@sephnews imagine writing some code that lowers worldwide electricity usage by 100%

8

0

139

Peter Baylies

@pbaylies

2 years

@AlphaSignalAI The word is "announced", not "released".

10

2

145

Peter Baylies

@pbaylies

3 years

StyleGAN2 projector modifications, with CLIP, addl. losses, and other cool things... this time as a gist, maybe eventually I'll have enough stuff for a repo. Feel free to play with it! cc: @l4rz @dvsch @advadnoun @erinbeess @theshawwn

5

16

138

Peter Baylies

@pbaylies

2 years

@thehardmenpath I know because I wrote code to mask out the backgrounds by default, which is why they're blurry in the encoded images.

GitHub - pbaylies/stylegan-encoder: StyleGAN Encoder - converts real images to latent space

StyleGAN Encoder - converts real images to latent space - pbaylies/stylegan-encoder

github.com

1

132

Peter Baylies

@pbaylies

4 years

@eturner303 @jeremyphoward

1

10

136

Peter Baylies

@pbaylies

3 years

Finetuned some Real ESRGAN upscaling models on paintings, check it out, folks! cc: @Buntworthy @RiversHaveWings @advadnoun @Norod78 @theshawwn @erinbeess @aicrumb Models are here:

Peter Baylies

@pbaylies

3 years

These are getting way too good and quick, tbh. :)

5

4

44

8

26

135

Peter Baylies

@pbaylies

3 years

4

14

120

Peter Baylies

@pbaylies

4 years

@MichaelFriese10 @ak92501 @theshawwn Ok folks, *now* I have a WikiArt conditional model released and a repo up! Model here: Repo here: Thanks to @davidstap @robertluxemburg @Veqtor @theshawwn for their repos and their work in this space!

9

24

117

Peter Baylies

@pbaylies

2 years

These No Man's Sky Screenshots Do Not Exist #StableDiffusion2

5

10

116

Peter Baylies

@pbaylies

2 years

#StableDiffusion2

4

9

109

Peter Baylies

@pbaylies

3 years

How can you put a huge neural network into production and still have no one notice? Deploy it on Bing!

/MachineLearning

@slashML

3 years

Microsoft AI Researchers Introduce A Neural Network With 135 Billion Parameters And Deployed It On Bing To Improve Search Results

6

45

238

2

7

109

Peter Baylies

@pbaylies

4 months

Seconded, I can't thank @EMostaque enough for bringing me on at @StabilityAI . He's a real one, don't believe everything that you read.

David Marx

@DigThatData

4 months

2y ago, I released pytti-tools. hours l8r, @EMostaque reached out to recruit me as 1 of @StabilityAI 's 1st eng hires. Easily 1 of happiest, most validating days my life. The co was SNAFU, but Emad is a gr8 guy who does a lot of good for open source AI. 😢

4

5

127

2

6

107

Peter Baylies

@pbaylies

2 years

Ran some stable diffusion hyperparameter tests today... Horizontal - scale - 2 3 5 7.5 10 12.5 15 16 17.5 20 25 30 Vertical - steps - 10 25 50 75 100 150 200 250 500 1000

3

7

104

Peter Baylies

@pbaylies

3 years

Made a little repo with my training code and simple models for augmenting CLIP embeddings; sample notebook included for doing CLIP guided diffusion. Thanks to @RiversHaveWings @danielrussruss @sadly_existent

GitHub - pbaylies/Augmented_CLIP: Training simple models to predict CLIP image embeddings from text...

Training simple models to predict CLIP image embeddings from text embeddings, and vice versa. - pbaylies/Augmented_CLIP

github.com

3

14

102

Peter Baylies

@pbaylies

2 years

I feel like I definitely don't work this hard on my prompts. What do you think? 🤔

Suhail

@Suhail

2 years

Expert level prompt engineering going on here:

180

147

2K

12

3

97

Peter Baylies

@pbaylies

3 years

@chloellene

3

4

93

Peter Baylies

@pbaylies

4 years

@macloo @mycoliza Eh; last I checked, it was easy to completely break pip by [checks notes] running the command pip suggests, to upgrade pip

2

1

86

Peter Baylies

@pbaylies

5 years

Starting a new experiment...

9

6

83

Peter Baylies

@pbaylies

2 years

First tests of the new ESGD-M second order optimizer from @RiversHaveWings - working well for optimizing StyleGAN2 latents!

0

5

84

Peter Baylies

@pbaylies

3 years

Here's the top level - 64 clusters built from 256k latents in my WikiArt model. Any favorites?

Peter Baylies

@pbaylies

3 years

Experiments with clustering 256k latents... Parent cluster 19, and then sub-cluster 52, and two samples therein...

2

0

26

24

5

80

Peter Baylies

@pbaylies

2 years

Really enjoying the new CLIP models so far, playing around with ViT-L/14...

8

3

80

Peter Baylies

@pbaylies

2 years

4

3

74

Peter Baylies

@pbaylies

4 years

@eturner303 @jeremyphoward

3

5

76

Peter Baylies

@pbaylies

5 years

Ok StyleGAN peeps, I've got a repo up, and it's a fork of Puzer's stylegan encoder, but now with a ResNet encoder (pretrained model available!) and more loss functions for the encoder to speed things up; check it out! #StyleGAN cc: @aarlo @neurokinetikz @calamardh @roadrunning01

5

14

74

Peter Baylies

@pbaylies

3 months

Dude, you're such a dick. Can you not?

Reid Southen

@Rahll

3 months

@moonride303 @EMostaque @StabilityAI Oh it's fun to see that @pbaylies liked this comment, he supports data laundering.

9

23

280

43

3

74

Peter Baylies

@pbaylies

5 years

@KylePlantEmoji I tried coding on a whiteboard once, but it didn't compile.

1

3

73

Peter Baylies

@pbaylies

4 years

@mark_riedl Appropriate, given that GPT-3 is a model that you can't touch...

5

2

68

Peter Baylies

@pbaylies

2 years

Updated my guided diffusion notebook fork and my little models, quality should be much improved!

7

68

Peter Baylies

@pbaylies

4 years

@dril_gpt2

2

62

Peter Baylies

@pbaylies

3 years

@jbhuang0604 Yo dawg, I heard you like deep features, so I used the feature embeddings as a representation to construct a latent space embedding the feature representations, so you can interpolate while you extrapolate.

0

4

66

Peter Baylies

@pbaylies

2 years

Twisted Mentat Matt

@moridinamael

2 years

Tip: Use #stablediffusion img2img to slightly enhance your profile pictures

13

20

154

7

6

65

Peter Baylies

@pbaylies

3 years

Very cool!

Daniel Russell

@danielrussruss

3 years

Here's where you can find the "much faster and direct way" to visualize CLIP attention: from @hila_chefer @shir_gur

0

7

70

3

8

66

Peter Baylies

@pbaylies

2 years

Got my invite! Any suggestions?

23

1

62

Peter Baylies

@pbaylies

3 years

Back to the old WikiArt StyleGAN2 model - DeceiveD is definitely improving the quality of the conditional model for generation...

4

2

64

Peter Baylies

@pbaylies

3 years

ruDALLE, "Образный фантастический пейзаж, картина Ван Гога."

8

5

64

Peter Baylies

@pbaylies

4 years

@ShadowBankerCEO @WrrrdNrrrdGrrrl @davidgerard What it should mean, is that you document what the code is intended to do, for future readers - similar to making notes in the margins of a book

2

1

59

Peter Baylies

@pbaylies

3 years

Got my clustering experiment code cleaned up and put in a notebook; have fun, folks! cc: @l4rz @Buntworthy @theshawwn @arfafax @advadnoun @erinbeess Some example clustering (0 - 15 out of 64) and subcluster 0 below:

Peter Baylies

@pbaylies

3 years

Here's the top level - 64 clusters built from 256k latents in my WikiArt model. Any favorites?

24

5

80

5

11

62

Peter Baylies

@pbaylies

2 years

@minimaxir Definitely agree that negative prompting is key! First one is with my usual weird negative prompting techniques (and some latent mirroring). Added the aggressive prompting to the other ones!

3

2

58

Peter Baylies

@pbaylies

4 years

@AlbertoPupo84 Betty and Veronica

0

1

55

Peter Baylies

@pbaylies

10 years

Check out my Vampire costume! #ScaryStoriesIn5Words http://t.co/kmMxg41gUU

2

42

55

Peter Baylies

@pbaylies

3 years

ResNet is the JPEG of ML models, the workhorse that still works better than it has any right to, given its age (2015)

William Fedus

@LiamFedus

3 years

Don’t dismiss old methods too quickly. ResNets are still strong baselines when augmented with modern improvements. Thanks for the shout-out, @OriolVinyalsML

0

11

87

0

4

57

Peter Baylies

@pbaylies

2 years

@borisdayma

4

2

56

Peter Baylies

@pbaylies

4 years

CLIP + FFHQ attempts - Queen Helen of Troy? It's really fun watching StyleGAN2 attempt to generate crowns!

3

6

54

Peter Baylies

@pbaylies

2 years

What's the current count on Google's unreleased SOTA image models, is it three, or more? Also, if you never release it, was it ever really SOTA? 🤔

Mike Ma (AI x Finance)

@MMikeMMa

2 years

@arankomatsuzaki @GoogleAI 🔥🔥🔥

0

16

6

1

54

Peter Baylies

@pbaylies

4 years

@crashfrog @ShadowBankerCEO @WrrrdNrrrdGrrrl @davidgerard Here's the thing - code has bugs; knowing what the author intended is often more important in the future.

2

1

52

Peter Baylies

@pbaylies

2 years

@J_Hunter_M08 @conradhackett

1

3

52

Peter Baylies

@pbaylies

3 years

Thank you @eps696 - this is quite fun!

vadim epstein

@eps696

3 years

illustrip [text-to-video animation tool] is freely available on … now, in both local and colab versions. the latter also includes 3d depth processing by @deKxi 👇👇👇

23

71

391

1

2

50

Peter Baylies

@pbaylies

4 years

@mcclure111 @TheKaufmann It worked in Dracula, all you need is a coffin full of your home soil, right?

1

0

48

Peter Baylies

@pbaylies

4 years

@tesseralis Cursed.

0

2

50

Peter Baylies

@pbaylies

3 years

Codex knows what's up. cc: @lucidrains

3

0

49

Peter Baylies

@pbaylies

2 years

You are standing in an open field west of a white house, with a boarded front door. There is a small mailbox here. #dalle #dalle2

9

2

46

Peter Baylies

@pbaylies

2 years

Let's see who remembers what this is...

3DTOPO

@3DTOPO

2 years

@pbaylies @jxmnop Tired: GANs Wired: Diffusion with adversarial loss

2

3

23

4

5

46

Peter Baylies

@pbaylies

3 years

Just put up a repo for my unsupervised clustering of the CLIP image embeddings for LAION-400m GitHub: Labels: Original dataset:

GitHub - pbaylies/clustering-laion400m: Script and models for clustering LAION-400m CLIP embeddings.

Script and models for clustering LAION-400m CLIP embeddings. - pbaylies/clustering-laion400m

github.com

2

3

46

Peter Baylies

@pbaylies

4 years

Nice paper, and I'm not entirely saying that just because they mentioned my StyleGAN encoder repo in it...

AK

@_akhaliq

4 years

StyleGAN2 Distillation for Feed-forward Image Manipulation pdf: abs: github:

18

260

1K

3

5

48

Peter Baylies

@pbaylies

4 years

@theshawwn It's also not a dataset issue that's making Obama look more like Seth Meyers. Here's (1) original, aligned; (2) PULSE with GEOCROSS loss used (sharper, yet whiter); (3) PULSE without GEOCROSS loss used (less defined); and (4) using a non-PULSE encoder into StyleGAN...

10

16

47

Peter Baylies

@pbaylies

2 years

@seldo And remember Allo? That looked promising for a minute... The successor there was Messages. It was released along with Google Duo, which is being replaced by Google Meet. Looks like a 3 to 5 year cycle?

1

46

Peter Baylies

@pbaylies

2 years

Cool story, bro

6

45

Peter Baylies

@pbaylies

3 years

GANformer ( ) samples @ 640kimg cc: @dvsch

3

4

46

Peter Baylies

@pbaylies

5 years

@glagolista Artisinal, hand-crafted, locally sourced neural networks.

0

3

46

Peter Baylies

@pbaylies

4 years

Very cool!

0

9

44

Peter Baylies

@pbaylies

3 years

These are getting way too good and quick, tbh. :)

Roope Rainisto

@rainisto

3 years

@MademoisellleK @congralilso @danielrussruss At

1

3

24

5

4

44

Peter Baylies

@pbaylies

2 years

#stablediffusion

5

2

41

Peter Baylies

@pbaylies

2 years

@LimnDigital

3

2

40

Peter Baylies

@pbaylies

5 years

@donnacamos88 1+1=11

4

1

41

Peter Baylies

@pbaylies

2 years

Obligatory AI self-roasting thanks to @pharmapsychotic

Peter Baylies

@pbaylies

2 years

This is the future @advadnoun warned us about...

2

3

20

7

6

36

Peter Baylies

@pbaylies

2 years

@PhillipBurgart @NGrossman81 Yes, but what was Congress wearing at the time?

1

0

38

Peter Baylies

@pbaylies

3 years

@advadnoun

1

38

Peter Baylies

@pbaylies

3 years

@mcclure111 Microsoft, they made it so boring that we don't even notice, remember, or know that it still happens anymore...

1

0

38

Peter Baylies

@pbaylies

5 years

@sineadwilliamso Parental Supervised Networks

0

2

39

Peter Baylies

@pbaylies

3 years

This just in, you can guide StyleGAN2 / BigGAN generation with CLIP by optimizing latents, Film at 11. Cc: @l4rz @advadnoun @eps696 @pbaylies etc. etc.

Artsiom Sanakoyeu

@artsiom_s

3 years

🔥New DALL-E? Paint by Word 🔥 Edit a generated image by painting a mask atany location of the image and specifying any text description. Or generate a full image just based on textual input. 📝 1/

2

11

65

3

2

37

Peter Baylies

@pbaylies

3 years

@mcclure111

1

0

36

Peter Baylies

@pbaylies

2 years

@JanelleCShane Trick question, all Waffle Houses are effectively already in Mordor! I like "WAFFHOLE" though, seems very appropriate.

0

38

Peter Baylies

@pbaylies

3 years

I tested out my unsupervised clustering methods on the first million CLIP image embeddings in LAION ( ), and then labelled the clusters by hand, here's what I found... cc: @theshawwn

Clustering LAION, the first million images - added counts

Clustering LAION, the first million images - added counts - gist:671b15e76411597d6bf7556271c08ce1

gist.github.com

3

37

Peter Baylies

@pbaylies

2 years

@chaosgonewrong @ai_curio

1

4

35

Peter Baylies

@pbaylies

2 years

Running some more tests, latent thresholding of 7.5, varying CFG Scale and Steps. (currently running the other test to compare against 🙃)

2

34

Peter Baylies

@pbaylies

5 years

@Zergfriend And the obligatory training video...

1

9

33

Peter Baylies

@pbaylies

10 years

Symantec AntiVirus Corporate Edition | Threat Detected [ Delete ] [ Quarantine ] [ Shake It Off ] cc: @symantec @SwiftOnSecurity

0

20

35

Peter Baylies

@pbaylies

3 years

Gather 'round folks, I have a story to tell. Trigger warning: may contain controversial topics, such as True Love, NFTs, or Facebook being used positively. This is how Artificial Intelligence helped me reconnect with the wonderful woman who would be my wife, @CarmenLBullock .🧵

2

35

Peter Baylies

@pbaylies

3 years

Or, and stick with me here, it has learned to read words a little bit, which honestly is much more impressive.

0

1

34

Peter Baylies

@pbaylies

4 years

I've got to say, I'm very impressed by what this model can do, already!

AK

@_akhaliq

4 years

github:

0

1

24

1

5

35

Peter Baylies

@pbaylies

2 years

Oh look, someone did a bad job of indexing LAION instead of just using what was already out there:

Arthur Holland Michel

@WriteArthur

2 years

This site let's you search the giant database behind image-making AI systems like Stable Diffusion. It's supposed to be for artists to see if their art is in the data, but it also shows the sheer volume of NSFW/toxic stuff that's behind these AI tools.

9

120

294

0

6

35

Peter Baylies

@pbaylies

2 years

@TheNerdStation

1

2

34

Peter Baylies

@pbaylies

2 years

@mknejp Farm for food, probably.

0

34

Peter Baylies

@pbaylies

2 years

@advadnoun They're just afraid of getting De-Monet-ized...

2

0

33

Peter Baylies

@pbaylies

2 years

#stablediffusion

2

3

33

Peter Baylies

@pbaylies

6 months

1

32

Peter Baylies

@pbaylies

2 years

You tell me. Honestly if you're not looking at the actual art but instead asking who made it or how it was made, that's called "gatekeeping", not "art appreciation".

THUMIN

@Thuminnoo

2 years

Do you consider AI artworks as "real" art?

135

8

379

0

4

30

Peter Baylies

@pbaylies

2 years

No Man's Sky themed AI art, love to see it! Cc: @NoMansSky @TheSurvivalBob @JasonPlaysNMS @CarmenLBullock @EMostaque @RiversHaveWings

Alexander S

@devdef

2 years

A set of variations of "a highly detailed matte painting of a No Man's Sky Screenshot, a tall grass field, broken machinery, Simon Stalenhag , featured on Artstation." Iirc @simonstalenhag did the original box art for @NoMansSky so this is a perfect matchup #stablediffusion