# This is not an AI Art Podcast (Ep. 3)


## Intro

Welcome to episode three! This is your host, Doug Smith. This is Not An AI art podcast is a podcast about, well, AI ART – technology, community, and techniques. With a focus on stable diffusion, but all art tools are up for grabs, from the pencil on up, and including pay-to-play tools, like Midjourney. Less philosophy – more tire kicking. But if the philosophy gets in the way, we'll cover it.

But plenty of art theory!

Today we've got:

* State of the Art: Automatic1111, is the community healthy?
* Model madness model reviews: Impressionism model, Uncanny & DuchaitenAI Model
* "Bloods and crits": Art critique on 4 pieces
* Technique of the week: Kind of three this week -- one tip, one SHORT traditional technique
* My project update: so you can learn from my process
* ...And a few PSAs

Available on:
* [Spotify](https://open.spotify.com/show/4RxBUvcx71dnOr1e1oYmvV)
* [iHeartRadio](https://www.iheart.com/podcast/269-this-is-not-an-ai-art-podc-112887791/)
* [Google Podcasts](https://podcasts.google.com/feed/aHR0cHM6Ly9hbmNob3IuZm0vcy9kZWY2YmQwOC9wb2RjYXN0L3Jzcw)

Show notes are always included, the format is intended to be so that you don't have to be looking at your screen -- but the show notes have all the imagery and prompts and details on the processes we look at.

## PSA: Don't catch GPS

Hi this is Sy Greenbloom, president for the society for the prevention of GPS -- do you have a case of GPS?

GENERIC PORTRAIT SYNDROME.

Are all your AI art pieces coming out with the same lame generic faces with no dialog but lots of photo realism?

YOU CAN PREVENT GPS.

Improve your concepts. And start iterating on your art. More on that later!

GPS is a theme of this week's episode.

## Model Madness

### Impressionism Model

https://civitai.com/models/28068?modelVersionId=46814

1920s flappers, impressionism oil painting by mse
Negative prompt: bad_prompt_version2:0.8
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: -1, Size: 768x768, Model hash: 880c3f37f3, Model: impressionismOil_sd21

Shoot, these came out pretty good!

Apparently made off of the artist's work. [reddit thread](https://www.reddit.com/r/StableDiffusion/comments/12wpiw7/brushstrokes_oil_painting_model_oc/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

Tip: Don't use restore faces (it makes them too photographic, at least with whatever restoration I'm using)

I'm doing a lot of painterly pieces for my main project, and I can definitely see myself trying to run some of my concepts through this model.


### Edge of realism

[Reddit thread](https://www.reddit.com/r/StableDiffusion/comments/12xm1d6/edge_of_realism/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

https://civitai.com/models/21813/edge-of-realism

1920's flapper, looking at viewer, holi color festival, portrait, hyper detailed, detailed face, candid photo, POV, by lee jeffries, nikon d850, film stock photograph ,4 kodak portra 400 ,camera f1.6 lens ,rich colors ,hyper realistic ,lifelike texture, dramatic lighting , cinestill 800
Negative prompt: bad_prompt_version2:0.8,  bad, jpeg artifacts, low res, bad lighting, deformed, mutated, black and white, monochromatic, comic, bad anatomy, bad hands, cropped, 1girl, 3d rendering
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: -1, Face restoration: CodeFormer, Size: 768x768, Model hash: c29496f597, Model: edgeOfRealism_eorV20BakedVAE

I definitely see the "uncanny valley"

https://en.wikipedia.org/wiki/Uncanny_valley

That's the like -- if it's too close to real but not real, it looks... dead or zombie-ish. I think of it as the like "creepy doll" effect, like weird creepy dolls have that uncanny valley thing to me.

I'm not exactly sure it's for me, but it is really neat, and I think there might be something to play with here in terms of "avoiding GPS"


### DucHaitenAIart Model

https://huggingface.co/DucHaiten/DucHaitenAIart/tree/main

I saw it from this [reddit thread](https://www.reddit.com/r/StableDiffusion/comments/12xexsf/duchaitenai_model_sometimes_just_blows_my_mind/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

The results from this are OUT OF THIS WORLD. This might become an everyday model for me. I absolutely love it.

There's part of me that thinks there's a possibility that the "lady in stone" we looked at last week came from this model.

1920s flapper, masterpiece, best quality, Lofi portrait, digital painting, HDR, Pixar style Painting by Joe Fenton, Stanley Artgerm, Tom Bagshaw, Tim Burton, sideways glance, foreshortening, extremely detailed 8K, high resolution, ultra quality, highly detail eyes, highly detail mouth, highly detailed face, perfect eyes, both eyes are the same, hd, 2k, 4k, 8k, 16k
Negative prompt: illustration, painting, cartoons, sketch, (worst quality:2), (low quality:2), (normal quality:2), lowres, bad anatomy, bad hands, ((monochrome)), ((grayscale)), collapsed eyeshadow, multiple eyeblows, vaginas in breasts, (cropped), oversaturated, extra limb, missing limbs, deformed hands, long neck, long body, imperfect, (bad hands), signature, watermark, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error
Steps: 20, Sampler: DPM++ 2S a Karras, CFG scale: 12, Seed: -1, Face restoration: CodeFormer, Size: 768x768, Model hash: c29496f597, Model: edgeOfRealism_eorV20BakedVAE, ENSD: 31337


## Cool resources

### Stable Diffusion Magic Prompt

https://huggingface.co/spaces/Gustavosta/MagicPrompt-Stable-Diffusion

Good for idea generation!

## LLaVa img2text

quick markdown from my notes when I installed it: https://hackmd.io/@dougbtv/Byl2jEEX2

I want to build some personal tooling to connect this and automatic1111.

## State of the art: Automatic1111

Automatic1111 is amazing for a number of reasons -- but the first in my mind is that it's ecosystem oriented, and that's the kind of thing that's building community.

My day job is working with open source software, not only developing software, but also working with and maintaining communities of people. One [working group](https://github.com/k8snetworkplumbingwg/) I'm involved with maintains over 30 github repos.

Honestly it reminds me a lot of an operating system. And if you look [operating systems on wikipedia](https://en.wikipedia.org/wiki/Operating_system) -- it fits the mold

> An operating system (OS) is system software that manages computer hardware and software resources, and provides common services for computer programs.

The way that automatic1111 has extensions really reminds me of powerful ecosystem things that have gotten me really excited when they were relevant (and even today)

Like [Perl's CPAN](https://www.cpan.org/) in the late 90's, [Node.js NPM](https://www.npmjs.com/) in the early 2010's, and to a lesser extent [CNI for networking](https://github.com/containernetworking) in Kubernetes.

It's amazing that people are cranking out new and interesting things for automatic1111, both as extensions, but also as modifications to the core codebase.

But -- [the maintainer](https://github.com/AUTOMATIC1111) of automatic1111 webui hasn't merged a PR into main since late March. And no activity on their github since then, too.

There was some big merge to main branch around that time, and it caused a bunch of regressions (things that worked no longer work). Tons of buzz on reddit about it, lots of people getting burnt by adding "git pull" to their startup script

I wonder if this might've been a straw that broke the camel's back for having a single maintainer instead of a community of maintainers.

**Pro tip: Don't track main branch**

There's 1.9k issues. There's 132 outstanding PRs.

What happens next?

This is bad news for the community. And there's a natural progression that's going to happen:

* People will fork automatic1111 (make their own copy)
* The community will become partially fractured into kind of potentially balkanized zones

> May you live in interesting times

Well -- we do!

This will be really telling what happens next.

I really hope that the author of automatic1111 is involved in the next steps, clearly they had a vision for what this would look like, and they did an awesome job and it clearly has adoption.

My fear is that commercial tools take over the space. Having this community driven project breeds COOL EXPERIMENTS. Things happen quickly! Amateurs can brew up neat interesting things, it keeps ideas fresh.

## PSA: Your concepts are razor thin!

Hi this is Sy Greenbloom again, president of the society for thicc concepts.

Are your concepts razor thin? Are they thinner than wood that's been shaved with a japanese planer?

Are you stuck creating rehashes of shitty LAME pop culture iconography? Have you been spending all your time making the pope wear puffy jackets? Have you been remixing the lord of the rings with raves? Are you on your 10,000th generation of Bart Simpson as a

Your concepts might be RAZOR THIN.

YOU NEED TO GET A REAL CONCEPT THAT PUSHES SOMETHING NEW.

I'm going to sound like my professors -- but this stuff sucks.

Like Steve's comics (he's a comic book artist now) or Brian's tattoo art, they trashed it, but... They weren't wrong.

## Bloods and Crits

### Soldier Ladies, on [reddit](https://www.reddit.com/r/StableDiffusion/comments/12w8681/female_solider_in_the_winter_workflow_in_the/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)


### Red lipstick, [reddit](https://www.reddit.com/r/StableDiffusion/comments/12w5bth/girl_with_red_lipstick/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

No workflow, is this a lucky?

Great rendering.

Concept is week. Slight GPS, but the dowdy face is great.


### Swimming in summer, [reddit](https://www.reddit.com/r/StableDiffusion/comments/12xr7j1/swimming_in_summer_is_this_real_enough/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

I chose this because it relates to work I'll share later.

Artist got smoked by downvotes in the comments. Wrong audience for their ask. Oh well.

Needs work on details and the narrative while not totally devoid, could be improved.

(image click in for NSFW)

### Tears of the universe [reddit](https://www.reddit.com/r/StableDiffusion/comments/12qjbss/the_tears_of_the_universe_by_mary_blair_and_tan/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)

Way to go mixing concepts! Awesome outcome. The "CD Cover" factor really comes out, so neat.


## Tip of the week

Inpaint faces on group photos

Ok, so here's where I started, and it's kind of a good start. The gist is that I took a midjourney output, and started reprocessing it with img2img. I wanted the same face, but I wanted to re-stylize the piece.

You can start with MUCH worse faces.


Then, I fixed up pieces that are crappy, there's a wonky hand on the left, and there's a hand in front of a face on the right.

So now I've got some patched up stuff that looks like:


Now I'm going to modify my prompt to basically be:

(my subject here:1.1), and the rest of your prompt here.

So in this case, take out details of your subject

Let's say the prompt starts like this:

a handsome 30yo man in a fur coat and hat is dancing around a table with women in the background and a beautiful woman in a fur coat, David Teniers III, promotional image, a renaissance painting, kitsch movement

(which is about what I was using)

And then change it to just:

(a beautiful woman in the background:1.1), David Teniers III, promotional image, a renaissance painting, kitsch movement

Now in inpaint, go paint a single face in, and for "inpaint area" choose "only masked" -- so this will paint a subset of your photo. Keep it high res, I'm using 1024 here into a 1024 image, so it'll scale it down after it renders it in detail

Ok, now very subtle here, it's the woman's face to his right.


Now we've got a perfect face. Now I'll keep sending it to inpaint and repeating for all the faces or ports of faces that messed up. I don't have a lot to fix here.

Additionally, I added some (subtly colorful) props to the table.


## Traditional technique: Measure with your pencil

Youtube tutorial:

https://www.youtube.com/watch?v=BavOGXGcR1o

## Technique of the week

Iterating your pieces

So I fed a tiny (very very low res) JPG web download to midjourney as an image prompt to try to capture a guy, and... I was happy with the results in terms of idea.

Here's what I got out of Midjourney, the idea's good. But it also gave me a "crappy web download" look to it, haha. Ok, that's not going to work. It's... too accurate haha.


I fed that into control net, using canny.

Here's some the initial image I picked from it:


And then I inpainted.

I was trying to go for a kind of "hand tinted early photography" look, which fits a style of photography that I'm working towards (and building a LoRA for)

* I painted over the canoe in blue.
* I changed the beard to look like the reference I was using
* I also fixed the hands.
* I added slats to the canoe, because that's accurate to the type of canoe used for the period and place I'm depecting
* I also did the lettering by hand

I wound up with something like this:


But it's got some problems to me. The light blue is like... too pastel and unique in the photo. The palette isn't balanced.

So, while I thought I had a finished piece, I didn't. I went to go look at my work, and the blue is too stark, and it doesn't relate to anything else in the palette. I went backwards in the workflow and changed to a darker canoe, and I added some blue in the sky to use related colors elsewhere in the piece.


I also like that the highest contrast areas are in portions that are the closest to the viewer in 3-space. Like the boots, knees, and hands. This is really good for the depth of the image.

THe composition isn't perfect, but I kind of like the triangles across the pieces between the boat in the foreground, and the land in the background, it gets your eye to move around the piece.

I'm still not happy with something about it. Now it's too dark. It especially looks too dark and doesn't stand out among other thumbnails. So, I feel like something is wrong with it.

I tried sepia toning it in an image editor and using auto levels.


I'm more happy with it. It's plausible I could do another round of colorization from here.

We have some additional contrast in places that brings our subject, the guide, into the foreground. Like, his face and the background are more distinct, and it highlights the subject better.

## Project update

On my main project -- I'm still trying to carve out the time, but I'm trying to build a tool to look at the Smithsonian Open Access, that is -- to download images by a query and then to have them be processed to use in training, especially in a LoRA

https://www.si.edu/openaccess

So, not my primary piece, but, somewhat inspired by it. I'm working on a series of "bathers".

I have a number of goals with it:

* I love figures, so I want something figural I can keep coming back to
* I love drapery, and I love water, and I want that water / drapery / figure.
* I want to explore more with light on the figure
* I want some consistency
* I love the kind of "luxury" of bathers in art
* Also my current main project has a kind of luxury and vacation vibes, and this plays into a similar feel, so it's relevant to me.
* When bathers were popular most recently, like, in the impressionism era, bathing was still a luxury.
* Heck, bathing is STILL a luxury. We just take showers.
* My favorite bathers are the Degas bathers (but this isn't based on them)
* It's also a sort of tongue-in-cheek commentary on the "hot babe pieces" that we're seeing in the ai art space. I want to get in on the fun, but I also want to poke fun at it in my own way
* I really want to play with fantastic poses, romanticism
* I also want to play with body type more.