AI Art – Let's Build Robotics with Shadow

A1111

Automatic1111 (A1111) is the go-to StableDiffusion (SD) image generation web interface for the vast majority of tutorials out there. While it’s not the easiest SD WebUI, A1111 is approachable by patient AI noobs and EasyDiffusion graduates alike. It exposes a bit many controls by default, but it packs enough power to keep a SD adept busy for a while. I also found Forge, an A1111 fork reportedly having extra features, bug fixes, grudging Linux support, and needs a virtual environment. At the top, I found ComfyUI, which lets you design and share custom workflows.

As a warm up exercise, I found a SonicDiffusion, a SD model geared for Sonic characters, generated a bunch of Ms. T portraits, and saved my favorites. Talking with my sister, I began cherry picking for details the model doesn’t control for, such as “cyclopse” designs where the eyes join at the whites vs. separate eyes (Hedgehogs are usually cyclopses, but not in the live action movies). SonicDiffusion –to my knowledge– lacks a keyword to force this distinction. Eventually, my expectations outpaced my ability to prompt, and I had to move on.

ControlNet

A major contribution to A1111’s versatility is its ecosystem of extensions. Of interest this week is ControlNet, a tool to include visual data in a StableDiffusion prompt for precision results. As of writing, I’m looking at 21 controller types – each needing a model to work. I downloaded the ones for Canny, Depth, and OpenPose to get started.

My first thought was to use an Xbox One Kinect (AKA Kinect v2) I bought from someone in my area a few Thanksgivings ago. If it works, I can easily pose for ControlNet. Long story short: I spent a couple days either last week or the week before tossing code back and forth with a self-hosted AI chatbot in SillyTavern with no dice. The open source Linux driver for the Kinect v2 just isn’t maintained for Ubuntu 22.04 and distros built on it. I couldn’t even get it to turn on its infrared LED’s (visible by my PinePhone’s camera) because of broken linkages in the header files or something. Pro tip: Don’t argue with a delusional LLM unless you can straighten it out in a reply or two. On the plus side, the AI did help me approach this job where I’d expect to have taken weeks to months without it. If/when I return, I expect to bodge it with Podman, but I may need to update the driver anyway if the kernel matters.

Even if I did get the Kinect to work, I doubt it would have been the miracle I was hoping for. Sonic style characters (Mobians) have different proportions than humans – most notably everything from the shoulders up. I ended up finding an embedding for making turnaround/character sheets, but it was again trained to make humans and I got inconsistent results compared to before. I did find a turnaround template for chibi characters that gave me OK-ish results running it through Canny, but Ms. T kept generating facing the wrong way.

On another session, I decided to try making Ms. T up in Sonic Forces. I installed it (ProtonDB: Platinum) and loaded my 100% save. I put Ms. T on a white background in GIMP and gave it to ControlNet. Unsurprisingly, OpenPose is not a Sonic fan. It’s trained on human data (now with animals!), but a cartoon kept returning blank outputs until I used a preprocessor called dw_openpose_full, which –while it still doesn’t like cartoon animal people– did cooperate on Ms. T’s right hand. Most every node else I dragged into place manually. I then demonstrated an ability to pose her left leg.

Character Sheet

From there, I opened OBS to record an .MP4 file. I used FFmpg to convert to .gif and loaded it in GIMP to… my computer slowed to a crawl, but it did comply without a major crash. I tried to crop and delete removed pixels… another slowdown, GIMP crashed. I adjusted OBS to record just my region of interest. 500+ frames was still a no-fly when each layer only has the changes from the last. I found options to record as .gif and record as slowly as I want. I then separated out my frames with FFmpg, making sure to have a new directory:

ffmpeg -i fileName.gif -vf fps=1 frame_%04d.jpg

I chose ten frames and arranged them in a 5×2 grid in GIMP. I then manually aligned OpenPose skeletons for each and sent that off to ControlNet. Immediately, my results improved. I got another big boost by using my grid of .gif frames, but in both cases Ms. T kept eyes and feet facing toward the viewer – even when her skeleton was pointed the other way. My next thought was to clean up the background on the grid, but compression artifacts got in the way.

Start over. I made a new character with joint visibility and background removal in mind. She looked ridiculous running through a level, but I got her knee placement by moving diagonal toward the camera and jumping. I then put eight new screenshots in a grid. Select-by-color had the background cleared in a minute. I then used Canny for silhouettes intending to reinforce OpenPose. I still got characters generating the wrong way.

Good Morning from my Robotics Lab! This is Shadow_8472 and today I am spending a week with Stable Diffusion to improve my skills at it. Let’s get started!

The science of AI art goes back to around the time complete CPU’s were first integrated into a single computer chip in the late 60’s/early 70’s. At least a couple waves of AI craze came and went, but on August 22, 2022, Stable Diffusion was released as free and open source software.

In the year since, Stable Diffusion has proven to be quite the disruptive technology. I’ve never had the cash to commission an online artist, but with a little effort, a decent amount of patience, and only an ounce of experience, I’ve gotten subjectively better results than commissioned works posted by low-end digital artists. I feel sorry for the people losing their dream jobs to machines, but at the same time this is a frontier I can have fun exploring.

One Week of Study

I’m setting myself a goal of spending two hours dedicated to learning Stable Diffusion every day this week. We’ll see what happens.

Monday

We won’t talk about what didn’t happen on Monday.

Tuesday

I finally started researching for this topic after midnight. I started up Easy Diffusion, an intuitive webUI for Stable Diffusion, generated a number of images with a project for my sister in mind.

I ended up looking up tips and tutorials. Looks like the hot-shot web UI these days is Automatic1111. It has more options, but is proportionally harder to use. I might try it later in the week. Otherwise, most of my time actually working today was writing the introduction.

Wednesday

Easy Diffusion is definitely the way to so if all you’re looking to do is goof around, because that is exactly what I did. So far as I can tell, I am at the exact bottom of graphics cards that can do this. I’m finding it of use to go smaller for faster feedback while learning to prompt. Conclusion: img2img has a tendency to muddle things.

Still, the draw of potentially more powerful techniques is calling. I found a piece of software called Stability Matrix, which supports a number of web UI’s – including Automatic1111, which every Stable Diffusion tutorial out there tends to go after. I ran into trouble with its integrated Python while setting it up (portable, in the end). I’m hoping I can replace it with a later version tomorrow.

Thursday

I switched approach from last night and did an online search for my error:

error while loading shared libraries: libcrypt.so.1: cannot open shared object file: No such file or directory

Multiple results pointed from people trying Python projects on Arch family systems like the one I’m on. One source from December 2022 recommended a multi-step process involving the AUR. I figured riffling through the project’s GitHub issues was worth a shot – to report it if nothing else. I searched for ‘libcrypt.so.1’, and the fix was to install libxcrypt-compat; I found it in the more trusted pacman repository [1].

AUR: Arch User Repository

I installed Automatic1111 using Stability Matrix and loaded it up. My first impression when compared to Easy Diffusion: Wall of controls. Easy is easy in both the setup AND the relatively intuitive control scheme, but it seemingly doesn’t support a lot of the tools I’ve seen and want to learn.

Per tradition, I made a photo of an astronaut riding a horse. It was a flop, but I got an image nonetheless. Its immediate followup didn’t finish when I told it to fix faces and I ran out of vRAM memory on my graphics card (to be fair, I didn’t have next to everything closed).

Sabbath starts tomorrow, and I’ve been writing these mostly late at night. I can tell I’m not likely to meet my time goal of a couple hours every day, but I feel getting to this step is a major accomplishment. Word count says 700+ words, so I could end it here and feel fine about it. I’ll see what happens. I want to find the control that tells it my graphics card is barely up to this stuff.

Friday

Time to start optimizing! For cotext, I’m on an NVIDIA graphics card with 4GB of vRAM, which is enough to get a feel for the software if you have a minute or two of patients per image, but having more would be prudent. After trying a couple online videos, I found AUTOMATIC1111’s GitHub had a list of optimizations [2] I’ll be listing as –flags to the COMMANDLINE_ARGS variable in my start script. I don’t have time this evening for a full test, but perhaps tomorrow night or Sunday I can do some benchmarking.

vRAM: Video RAM (Random Access Memory) *So glad to have finally looked this one up!*

xformers

For NVIDIA cards, we have a library xformers. It speeds up image generation and lowers vRAM usage, but at the cost of consistency, which may not be a bad thing depending on the situation.

opt-split-attention/opt-sub-quad-attention/opt-split-attention-v1

A “black magic” optimization that should be automatically handled. I’ll be selecting one via the webUI, though.

medvram/lowvram

This optimization breaks up the model to accommodate lesser graphics cards. The smaller the pieces though, the more time it will need to swap pieces out. Side note, but I believe it’s MEDvram as in MEDium as opposed to the naive pronunciation I heard with MEDvram as in MEDical.

opt-channelslast

Some procedures are exparimental optimization that is literally unknown if it’s worth it at this time. I’m skipping it.

Saturday Night

I took it off

Sunday

I joined my father on shopping trip and we ran out of gas at a car wash. By the time I sat down to work on Stable Diffusion, I wasn’t up to much more than an unguided self-tour of the settings. I don’t know what most of the categories are supposed to do! I’ll look each one up in time.

Monday Morning

As usual in recent months, I spend a while writing the Takeaway and Final Question, dressing up the citations, and copying everything out of LibreOffice and into WordPress for publication at noon.

Takeaway

Progress! It might not be what I expected this week, but I’m still satisfied that I have something to show off. The point I’m at is to get to the same place as I was with Easy Diffusion before looking up the toys I came to Automatic1111 for.

As one final note, this week is also the anniversary of this blog. It caused a bit of a delay in getting this post scheduled by noon, but that would make it the third instance I can remember of a late post in twice as many years. I feel bad about it, but at the same time, it’s still a decent track record.

Final Question

Do you have a favorite interface for using Stable Diffusion?

[1] PresTrembleyIIIEsq, et. all, “SD.Next / ComfyUI Install: Unexpected Error #54,” github.com, July 30, 2023. [Online]. https://github.com/LykosAI/StabilityMatrix/issues/54. [Accessed Aug. 8, 2023].

[2] AUTOMATIC1111, “Optimizations” github.com, August, 2023. [Online]. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations. [Accessed Aug. 8, 2023].

Tag: AI Art

Self-Hosted AI Consistent Characters: Part 1