## Intro

Welcome to episode sixteeen! 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: A few lora LoRAs for SDXL
* Bloods and crits: Just one
* Technique of the week: Doing some "image overlay" using the `/remix` feature of Midjourney.
* Updates on my project: So you can learn about it, doing some "southern gothic" kind of stuff.

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.

## News

Midjourney finally has inpainting! [From reddit thread](https://www.reddit.com/r/midjourney/comments/15xn5e2/mj_now_has_an_inpainting_system_called_vary_region/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)

I'd say... I'm glad it has it. But... it's such a roll of the dice that it's not nearly as powerful as using SD.

So if you're already an SD user, and an MJ user, just stick with SD usually. Althought you might try it on some stuff where you just need a "area re-roll".

[3d representations found in the weights for stable diffusion](https://www.reddit.com/r/StableDiffusion/comments/15wvz2a/researchers_discover_that_stable_diffusion_v1/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)

Interesting because it's not trained on that.

Similar kind of thing with GPT-4 having visual capability, [arxiv](https://arxiv.org/pdf/2304.10592.pdf#:~:text=For%20example%2C%20GPT%2D4%20can,based%20on%20handwritten%20text%20instructions.).

## Model Madness

### More details LoRA for SDXL

Available @ https://civitai.com/models/126343
From [this reddit thread](https://www.reddit.com/r/StableDiffusion/comments/15niij6/sdxl_better_picture_more_details_lora_download_on/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)

Images from it are rather stunning, but, also look a bit oversaturated.



### SDXL Pixel Art LoRA

From [reddit thread](https://www.reddit.com/r/StableDiffusion/comments/15otedd/some_early_samples_of_a_sdxl_high_detail_pixel/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)
[On civitai](https://civitai.com/models/120096/pixel-art-xl)

,%20%20(minimalist_1.3),%20_lora_pixelbuildings128-v2_1.2_,.jpeg)

### Negative Prompt LoRA (SDXL)

* [On reddit](https://www.reddit.com/r/StableDiffusion/comments/15r5k3i/i_just_released_a_lora_for_stable_diffusion_xl_10/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)
* [On huggingface](https://huggingface.co/minimaxir/sdxl-wrong-lora)

I have to admit I don't "get it" with negatives in SDXL yet, seems like you barely need them. But this seems to have a good effect.


### Other resources

* Variation seed to make variations like MJ
* [Article on civitai](https://civitai.com/articles/1688/basic-things-you-might-not-know-do-you-know-sd-has-vary-function-just-like-midjourney)
* There's a new sampler, DPM++ 3M SDE GPU
* I really like DPM++ 2M Karras, so I want to give it a try
* [From this reddit thread](https://www.reddit.com/r/StableDiffusion/comments/15qw3ri/demo_new_sampler_dpm_3m_sde_gpu_and_sgm_uniform/)
* The hidden Kohya SS docs
* Some are so arcane, and I'm often guessing! it's nice to get the pointer
* [From reddit thread](https://www.reddit.com/r/StableDiffusion/comments/15qw8yk/kohya_documentation_for_every_lora_training_option/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)
* [On github](https://github.com/bmaltais/kohya_ss/blob/master/docs/LoRA/options.md)

## Bloods and crits

### Prayer

OMG this is amazing. Click into the link and watch the video.

I really like that they used

Also, it's worth noting that almost everything came out super well because the artist had control of this tool from start to finish.

Sure, we like the "gambling aspect" of generations, but, being at the helm of the ship is a recipe for success.

* From [reddit](https://www.reddit.com/r/StableDiffusion/comments/15q324o/concept_art_timelapse_5_hrs_in_5_min/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)
* [final piece](https://www.behance.net/gallery/177584881/Prayer)


## Technique of the week

Subtle overlay effect with MJ.

Which honestly -- if you're not an MJ user, you can achieve the same idea by going img2img, changing the prompt, and playing with the denoise. Or, even moreso, you can try using the input image as a control net and playing with it there.

Inspired by [this reddit post](https://www.reddit.com/r/midjourney/comments/15ztvji/remix_subtle_to_get_an_overlay_effect/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=1)

1.	Type the command: /prefer remix
•	This sets a preference for remixing or modifying an image.
2.	Type the command: /imagine and prompt for a portrait.
•	It’s easier to choose a face with few details. Upscale the one you want.
3.	click the box: vary (subtle)
•	This box will only show if you’ve toggled remix in step 1
4.	A dialog box or prompt appears, asking you to describe what you want to overlay with. Type your description.
•	For example, you can say: “a rural landscape”

I started with this prompt and image:

portrait of a steampunk goth woman in the 1890s, by Nan Goldin


Then I tried to modify the prompt to:

steampunk clockwork, by Nan Goldin

Didn't work that well honestly. I think keeping `Nan Goldin` there still references a bunch of portrait photography.

If you're not familiar with Nan Goldin, definitely check her work out. [Google images search](https://www.google.com/search?sca_esv=559765737&sxsrf=AB5stBhR3F2nIVULubUu0N9huKdUyjRT2w:1692914213962&q=nan+goldin+photography&tbm=isch&source=lnms&sa=X&ved=2ahUKEwjrqe7ypPaAAxWRMlkFHYUeAE8Q0pQJegQIDRAB&biw=1920&bih=923)


So I tried changing the prompt ENTIRELY. Which is what I think you need to do...

flowers by John Singer Sargent


And the further you change it from the original image, the stronger the effect!

cyberpunk cityscape


## Update on my project

Funny how sometimes things that you're only medium on are hits with your audience. It's a good reminder that as much as you can be self critical, if you're making art for an audience, sometimes they can give you great feedback.

Been slacking on making LoRA's! I'm just short on time. And also playing with SDXL. Been thinking of trying to brew up a couple more LoRAs though. As soon as I can get ahead (again) with my main project.

I'm kinda waiting for some community movement on SDXL, because I'm happy with the workflow I'm using which is mostly MJ for initial generations, and then SD for inpainting and iteration -- and especially: control net.

Found someone who's got a regionally specific instagram account and they're doing a "southern gothic" style for the region I'm working on for my historical fiction project, and I'm PUMPED on it. It's so much fun!