Author: Andrew_Krill

Why?

There’s been a lot of negativity towards “AI Art”, especially from the “traditional” art community.

Most people think of “AI Art” as a one-click one-and-done workflow. You type a prompt, hit “generate” a few times, and voila! This may be true in certain cases, but you’ll end up with some wonky results:

Sure you can fiddle with the prompts, click “generate” a hundred times until you get the perfect result.

I want to show that AI art can be worthwhile, and a useful tool, for bringing concept to image.

How?

You can get extremely granular control of image generation by using a few techniques, such as img2img in-painting, or manually seeding the image in your favorite image editing application. In this example, I wanted the output to be purely generated by AI. However, I’ve had great success by manually manipulating or painting low-res features, then in-painting them to completion.

Much like a painter’s workflow, i.e. sketch, add the base colors, shading, add detail, this will be a very iterative process.

1: The starting point

I knew exactly what I wanted the final image to look like. Since this was a remix/parody/commentary, I started with M.C. Escher’s famous work.

Maybe I will do a write-up on starting from scratch in the future.

2: Masking

The most important part of the process is making a good mask. This is easily accomplished in Photoshop.

Once you have a good mask, you’re almost there.

Hitting “generate” a hundred times can work, but you might not get the “feel” you want. Also, in my case, the generated hand was kind of wonky and definitely did not fit the style of the work.

3: Training with Dreambooth

To get the pencil style of the work, we can zoom in and crop out pieces of the drawing. Zoom in too far, you’ll only have noise. Don’t zoom in enough, the concept learned will be the hands and pencils in the drawing. Took 3 times to get this right.

I used Dreambooth with the standard recommended settings. ~100 steps per image. I used 24 images and got decent results. Here is a cat in this style:

I used a similar process to get the style of robotic hands that I wanted, as well as the abstract drawing of the network/schematic/math. Took about 30 images of robot hands, and about 100 images of network architectures from various papers on AI.

4: Generate an Output

I started with the sleeves using an in-paint mask. I used “original” as the masked content fill method, 30 steps with Euler a, low CFG scale, and a denoise strength of about 0.3. I wanted the AI to retain the rough shape of the sleeve, and keep the relative level of detail to be low, to match the sketched sleeve.

Using the second mask with the hand,

I was pretty happy with this result. However, there are a few imperfections. Like any art process, this is iterative. You won’t get perfect results first try.

Look at the fingers. Look at the pencil. The style could be a bit better as well.

5: Mask, Generate, Repeat

It also took some manual fixes. The clone and smudge tools in photoshop are a good way to reshape things. The AI will recognize the texture is supposed to be part of the object you’re filling in. Masking was used to retain the sharp point of the pencil. I had to manually straighten the pencil as well.

6: Upscale

I had the best luck using a 50/50 of SWINIR and ESRGAN. Using one or the other resulted in some parts of the piece being extremely sharp, others being blurred or distorted.

7: Done

I’ve made a lot of projects over the years, and I really haven’t captured them in a good, meaningful way. I’ve been intending to write up these projects for years now, but I just don’t have time time to dedicate a whole post and write-up to each project. Instead, I will post pictures of each project and do a mini-writeup for each project.

Steak Brand

A friend of a friend had a film project, and they needed a good way to brand a steak with the words “Flambe”, the title of the film. I learned a ton from this project, it was one of my first metal working projects ever.

Lessons learned: steak branding should be done with steel, super glue shouldn’t hold items that are HOT, and the brand should have been reversed.

• 

• 

2020 has been a crazy year. Pandemic, Protests, Murder Hornets, and no end in sight!

I had joined a Milwaukee activism facebook group for donating supplies. One member had a megaphone to donate, but lost the charger!

I offered some of my time to fix it- either find/make a new charger, or find a way to make it run off of a common type of battery available at Walmart or Walgreens.

First I checked online- is this battery common or cheap? How about the charger?

Nope! $100 for a charger! Not worth just buying one. The battery comes for around $100 as well, I could save $200 if I can fix this!

The first challenge was finding out the voltage and pin-out of the battery. That is, which is (+) and which is (-). It is a proprietary battery pack, and I was not able to find any information on it online.

So, I took it apart and took a look at the 4 battery pins on the circuit board.

Right away I saw that one pin was connected to the “ground plane“, I marked this in BLUE. Circuit boards start with a flat sheet of copper and etch away around wires and connections. In this case, most of the sheet has been left intact to connect all of the ground connections (-). In some boards there are also “power planes”, where the whole sheet of copper is left behind and connected to power pins.

Another way to find the (-) ground pin, I used my multimeter with the ohm (Ω) mode to measure resistance, and measured between each pin and any screw. The meter read 0.1Ω on this one pin, so this was it!

Next, I noticed that the two battery pins highlighted in YELLOW both connected to capacitors. To me this indicated that they were the communication lines, for the megaphone to check on battery status.

This left RED as the only pin left for (+). Unfortunately there was no way to check where this really connected, and I couldn’t remove the circuit board from the enclosure, as it was glued into place from the other side.

To test my hypothesis on which pins were (+) and (-), I ended up charging the battery slightly. It managed to hold a slight charge, and I was able to measure it with my multimeter on the VOLTAGE setting.

The side of the battery says “11.1V NOM”, which means the battery fully charged is usually around 11V. Batteries usually have a higher voltage when they are freshly charged, so this could be around 13V fully charged! And since batteries lose charge as you use them, this voltage can drop considerably. Half-way through a use, it may measure down to even 9V! So whichever battery solution I find, it had to be within 9-13V.

I then used a power supply to try powering it up. I used a lab supply, but anyone could use a “wall wart” supply, so long as the voltage matches.

9V worked! So I went ahead and ordered some 9V battery holders from amazon:

And once it arrived I tried it out… It powered on! But when I keyed the siren key it would quickly die!

Shoot! 9V Batteries have a high internal resistance. That is, drawing lots of current, the voltage will drop considerably! And, a megaphone draws a lot of power!

While D-Cell batteries may provide enough current, but they’re only 1.5V! To get 9V I would need to stack 6 in series (“in a row”)! To get 12V I would need 8! Plus these batteries are HUGE and can be heavy!

So, I thought, what about a lithium battery? I have some very bright LED flashlights, they use lithium 123A batteries, available at your typical Walgreens or Walmart. These run at 3V and can supply considerable amounts of current!

These actually fit my requirements perfectly- Small enough to fit in the battery slot, powerful enough to power the speaker and amplifier, and I should only need 4 (4x3V = 12V!). So, I went ahead and ordered some 123A battery holders:

So, I went ahead and soldered 4 of these together, then ran wire down to the battery pins and soldered them there.

Then, I used tape to secure everything.

It worked perfect! I finished taping it up, then lent it to some activist friends to use.

• 

I was working on a personal project in the UWM Makerspace the other day, when Chris Beimborn, the Outreach Coordinater of the College of Engineering and Applied Science (CEAS), as well as the leader of a program called EnQuest, an Engineering program for high school age girls.

Over the past semester, they have been working on assembling Solar Panel stands and wiring up some cellphone charger boards for a project for a village in Guatemala. This mountainous village has no electricity, since the power lines were cut during a civil war. However, they still have a need for electricity, and cell phones for internet access. The area has cell phone reception, just no electricity. This is their means for communication with the world.

So, EnQuest was assembling these boards for the solar phone chargers, so villagers wouldn’t have to walk to the next village over to charge their phones. These chargers were then to be brought over by the UWM chapter of Engineering without Borders.

So, Chris comes to me asking if I could take a look at some boards. The boards were shipping out the next day, about about 10 of them weren’t working properly, as well as 2 solar panel assemblies. The SMD work was done by a group of volunteers who weren’t able to repair them in time, so I decided to take a look.

First I started with the datasheet of the only IC on the board, the TPS5516. Looking at the recommended schematic, I drew it out on a whiteboard. 

I then compared it to the printed circuit board I had on my desk, and annotated the schematic. I probed around with a multimeter. On some of the boards I was getting some odd readings, the V_Sense and V_Reg pins were reading the 5V but then the V_Out was reading 2.9V. The PG pin wasn’t on, so the power indicator light was off. PG ended up standing for “Power Good”

Probing around more, all of the other components were reading fine. I then decided to look under our microscope to see if I could find anything that was wrong with the soldering job. Immediately I was greeted with two mistakes notorious with SMD work.

So, pins were bridged on just about every faulty board. In addition, on half of these boards, the capacitor that controls the oscillation for the buck converter wasn’t even making contact!

After cleaning up the solder jobs and cleaning the flux off the board with isopropyl alcohol,  I was able to restore 9 out of 10 boards back to working order! The 10th board had a trace that was ripped off from thermal stress when the high-schoolers were soldering the large inductors on.

We were then able to package and ship them out, just in time for the Engineering without Borders team to take them to the airport.

So, the other week I was in the hospital for extreme abdominal pain. Turns out, after a CAT scan, I found out I had kidney stones. Worst pain I’ve ever felt.

On the plus side, I was able to ask the CT techs for the DICOM files! DICOM is a protocol that stands for Digital Imaging and Communications in Medicine, and this is what the CAT scan machine spits out.

What a CAT scan machine does is take an X-ray while moving the patient (me) and rotating the x-ray aperture (emitter) and corresponding sensors. The procedure took a couple of minutes, and the result was thousands of individual x-ray image “slices” from all different angles. Using clever maths, it’s possible to use those shots from multiple angles to determine the density of whatever is being imaged. In this case, it was my abdomen and kidneys. The darker areas shows are areas that did not absorb much radiation: skin, muscle, fat, etc. The lighter areas are areas that bounced back much of the radiation, namely my organs and bones.

The CT tech’s computer takes all of these individual slices, stitches them together, and creates a 3D model that they can analyze, and view in both 3D and 2D. The 2D slices generally correspond to the 3 planes of the body, Sagittal, Coronal, and Transverse, planes. Through these planes, they can pan back and forth through the body.

With those sliders, it’s possible to pan back and forth, through the body, looking at all sorts of different layers.

Now, the files the techs gave me were only 2D, so I used a program called 3D slicer. It can be found here: https://www.slicer.org/ . It is a very powerful program, allowing you to identify, or trace, over the individual slices. The end result is a 3D composite made from the individual slices. Here is what the output from 3d slicer looked like:

Note that there are a lot of lumps, ridges, and points, that are due to artifacts from the slicing process. I then put the raw .obj files and opened them up in Z-Brush. Z-Brush is a powerful and intuitive sculpting and modeling program, usually used for digital artists. With it, I was able to remove the many artifacts and smooth the kidney out to what it should look like:

I then opened up the .obj file in Slic3r (not to be confused with 3d slicer), added support material, and got the file ready to be printed. I exported it as gcode and got it ready to print!

 

Unfortunately I haven’t had a chance to actually print the kidneys, I’ve been too busy with work and school.

I plan to stop at Milwaukee Makespace this Tuesday, and I will upload the pictures of it as soon as the print is done!

Welcome to my website! On here I will be keeping a blog, adding content, keeping a portfolio, and more!

Check me out on GitHub: https://github.com/AndrewKrill