# This is not an AI Art Podcast (Ep. 5)
## Intro
Welcome to episode five! 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:
* Model madness model reviews: 3 LoRAs, and 2 kind of teasers for models
* "Bloods and crits": Art critique on three pieces
* Technique of the week: Compositional control. I'm always harping on it, I'll show you how to fix it.
* My project update: so you can learn from my process
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 and include all the visuals, prompts and technique examples, 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: Subject Placement Awareness League (SPAL).
Hi, this is Sy Greenbloom, president of the Subject Placement Awareness League (SPAL).
Are all of your compositions just a subject centered statistically accurately in a frame?
Do your compositions have so little movement as to keep the eye centered on the frame and keep the viewers attention for less than 20 milliseconds?
You might have a subject placement problem that's leading to LAME composition.
## Model madness
### "Detail Tweaker" LoRA
From [this reddit thread](https://www.reddit.com/r/StableDiffusion/comments/138easo/ive_made_a_detail_adder_loraive_made_a_detail/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)
https://civitai.com/models/58390/detail-tweaker-lora-lora
Here's the prompt, using Analog Madness 4.0
```
1920s flapper in a busy nightclub at night, bokeh, depth of field, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt <lora:add_detail:1.0>
```
### Crazy Expressions LoRA
https://civitai.com/models/5891/crazy-expressions
```
1920s flapper in a busy nightclub at night, constricted pupils, crazy face, bokeh, depth of field, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt <lora:Crazy_ExpressionsV2:1>
Negative prompt: bad_prompt_version2:0.8, ((((big hands, un-detailed skin, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing)))), (((ugly mouth, ugly eyes, missing teeth, crooked teeth, close up, cropped, out of frame))), worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, (skinny:1.3)
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: -1, Face restoration: CodeFormer, Size: 768x768, Model hash: 0b914c246e, Model: analogMadness_v40
```
```
1920s flapper dancing in a busy nightclub at night, constricted pupils, crazy smile, bokeh, depth of field, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt <lora:Crazy_ExpressionsV2:1>
```
### Teaser: FusionAI Model
From this [reddit thread](https://www.reddit.com/r/StableDiffusion/comments/138zreq/testers_for_the_fusionai_model_sd_21/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)
This looks crazy.
### Just a preview: Baldur's Gate Model
From [this reddit thread](https://www.reddit.com/r/StableDiffusion/comments/13977si/trained_a_model_on_a_bunch_of_baldurs_gate_maps/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)
Yep, this is going to be good for gaming. Looks so classic.
Up to you what you're going to train with, and all of that aside: Using something that you can readily get a lot of images from (like a video game) and then using that as training material is a gold mine of training data.
### Warm minimalism LoRA
From [this reddit thread](https://www.reddit.com/r/StableDiffusion/comments/1373diw/warm_minimalism_lora/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button)
Download @ https://dreamscapeapp.com/models/TIMD06c8ade9925fd95f8e72cdeb7e364c
```
a minimalist study in a western lodge, hardwood desk, globe, wooden chair, curio, exposed beams, volumetric fog, curtains blocking window, sunlight, wooden walls, wooden shelves, laptop, lamp, in the style of xyz <lora:warm-minimalism:1>
Negative prompt: bad_prompt_version2:0.8
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2651505415, Face restoration: CodeFormer, Size: 768x768, Model hash: 0b914c246e, Model: analogMadness_v40
```
Looks like it might be overfit. See how we're getting almost all the same room design. And not a lot of my prompt is shining through.
### Bonus resources
Someone's dump of thousands of controlnet poses
https://civitai.com/models/58554
Apparently a failed business model, RIP. Thanks for the poses.
## Bloods and crits
### Miniature worlds [reddit](https://www.reddit.com/r/StableDiffusion/comments/13cx5up/im_addicted_to_creating_miniature_worlds_more/?utm_source=share&utm_medium=web2x&context=3)
This is a really fun idea and they include the workflow -- if you haven't created miniatures before, you gotta give it a shot, it's really satisfying somehow.
Most of the pieces have the subject centered in the frame. I actually picked one that JUST goes off the frame. And I think it helps with the look of depth.
Having them centered does look tiny, but what else could we do to make them look tiny? Potentially part-to-whole relationship showing something for scale. Here we have a pen tip, and that helps.
But overall, we can use techniques like we're talking about today to control the composition even further.
### Tropical places I'd love to visit [reddit](https://www.reddit.com/r/StableDiffusion/comments/13buaev/tropical_places_i_would_love_to_visit/)
These came out really well, and I notice they're using the detail tweaker LoRA. It's working really well for them, and the detail is crazy. Rendering quality is awesome, and the colorful sunny places.
The fantastic look of the place adds a lot of narrative.
There's some good repetition of form with the palm fronds. And there's good depth to the image overall. I'd love to see the repetition of form pushed a little further even. Also I wonder if the series could be made in such a way as to kind of tell a story in a number of frames.
Nothing detracting from the pieces, though.
### The Orient Express [reddit](https://www.reddit.com/r/StableDiffusion/comments/138jgk4/the_orient_express_7680_x_5120/)
I love the style!!! It's very creative and surreal. It's different. It's really thought provoking.
You guessed it, I'm going to point out that the negative space not being broken up on the outside of the frame is detracting from it. Granted, there's a lot of surrealism that has this like stark desert look with shit clearly in the frame. But there's a lot of opportunity to kind of get more movement of the eye.
Honestly if you just crop this and take off a little bit of the left side of the caravan thing, it makes the image pop a little more into 3 space.
### Lots of colors [reddit](https://www.reddit.com/r/StableDiffusion/comments/13duovp/lots_of_colors/)
Not my style, and the bubble-gum brightness is crazy, but it's really effective in a number of ways. I like the setting a lot, and the soda machine and store setting is really cool and adds to the story, and I think the narrative overall is rather good. Hands came out well!
The color pallette is REALLY working well, and the color work very well and there's like nothing out of place color-wise, which looks challenging with a piece like this. Artist shared that they inpainted it for hours. I get it.
I think the biggest detraction from the piece is the "twins" look. It's something that we see a lot when we've got multiple subjects is that we get twins / clones. If that was changed a bit, it'd really help.
Also there's a ton of symmetry going on here. I want something more to pull us away from it. The background is helping a little bit, and a few details, but it's so symmetrical that it's not moving my eye around enough to help see all the cool detail that's been added.
I think the heads are supposed to be big for the style. With normal human proportions their feet would be 4x times below the bottom of the image. But, those heads are too big. I'm not big on the big eyes and kitty ears because they look "too typical" and come off as stereotypical, but I'm probably the wrong audience. So I can't say it detracts from the piece.
Granted, I like academism and romanticism, so... Who am I to judge?
## Traditional technique
Step away from your work.
Both to see the mistakes, and to get a fresh pair of eyes when you come back. You tire yourself out.
## Technique of the week
Using controlnet to control design elements.
Composition, and repetition of form.
First I start with a < 5 minute drawing
I'm trying to create some movement of the eye with the negative space (the water), and I'm also trying to get repetition of form.
I load it in as a control net scribble, and I use a prompt similar to this:
```
top down view of a stony river stream, clear water, reflections on water, detailed stones, smooth stones, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt <lora:add_detail:1>
```
(This used analog madness, model, too)
I also used it as img2img with a super high denoise, like > 0.8
And I wound up with something like:
Another quick sketch, just looking for something that moves my eye around and is angular.
I came up with a prompt like this, and ran it through with the DucHaitenAI model...
```
cyberpunk server rack, rear panel wiring, detailed network switch, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt <lora:add_detail:1>
```
And I get something like this:
Sure. It's cool, but it's typically just a server rack in the middle of a frame.
But when I apply the scribble, I get something with much more movement.
And to pick one from it...
### And one based on a Rothko
A.K.A how to use €140 worth of art (and butcher it a little hah!)
So Rothko knows a thing or two more about color and composition than I do, he was able to achieve a lot with very little, I mean, he's an abstract minimalist.
So I used his 1951 piece, `No. 6 (Violet, Green and Red)` [from wikipedia](https://en.wikipedia.org/wiki/No._6_(Violet,_Green_and_Red)) as img2img (no control net)
Set denoise to 0.66, and I used this prompt with Analog Madness:
```
river landscape, RAW photo, 4k, 8k, UHD, film grain, analog style, Fujifilm 300 xt
```
## Bonus piece
I butchered his proportions, so he's probably rolling in his grave. But the for the purposes of controlling our compositions in stable diffusion.
I wound up being inspired by making a Rothko imitation, and made this piece, Rothko with a Rotko, in a landscape with the painting supposed to match the landscape.
I'm overall happy-ish with it. I've been doing a lot of pieces that are supposed to have this kind of hand-tinted photography look, and I'm so-so on that. But the composition is there. We get this kind of 1,2,3 subjects going from foreground to background -- dude, painting, mountains. Then we get these series of horizontal lines from the bottom up and into the background, stone wall, frame, painting, middle of painting, island, edge of lake, top of painting, top of frame. I really like that.
I first generated a portrait of Rothko in Midjourney, and then fit it to match my idea of having rothko with a painting that matches the landscape. So I took the MJ generation and then photobashed it.
I cut out Rotkho from it. And the painting from the MJ generation. And then made that transparent out of it, using a landscape that I borrowed.
I used that as a control net. I then generated a color dude. I cut that dude out and added it to the photobash, and then used that for img2img, keeping the control net
### Bonus technique
Tile upscale.
I'll let Sebastian tell you the details: https://www.youtube.com/watch?v=EmA0RwWv-os
Note: You don't need to check out the branch anymore, it's merged into main, so just install using the git URL.
## Project update
Working through recipes for installation of tools in my lab
* Stable Diffusion: Success!
* Everydream2 trainer: Success!
* Did my first models
* Oobabooga: failure.
Going through highs and lows with my project. Sometimes I feel like I'm getting super winners, and super losers, but it's my own perspective.
People are asking me on my social for my project about if it's AI, I'm just straight up about it, but I try to explain it with a professional and artist's bend. I also am bringing up that there's more to this art than it just being generational. Had an artist that's a photographer contact me asking for more detail.
Sign in with Wallet
Connect another wallet