Dane.Kouttron

[09.02.22]  Synthetic Art meets an Acrylic Painting Robot

Its September 2022, Ethereum switched to proof of stake, 'NICE' 30-series Nvidia GPU's went from 1k$ to 200$ and a wild frontier of synthetic 'AI' art was on the horizon. Lets explore Synthetic Diffusion and how it applies to robotic art.

Background What is Stable Diffusion (SD)? The goal of SD is to generate a digital image from natural language description, or text to picture. This is image synthesis not image lookup, so the process is way more intricate

Some up close images of the gadget

Steps and Guidance Two of the major variables that are immediately accessible to the user are Number of Inference Steps and Guidance Scale. Lets make an analogy, solving our text to picture is just like having an exam question. In the simplest form STEPS is proportional to how long the system has to find a solution to a problem, akin to how long you get to look at a question during an exam. Just like in an exam, there's a best amount of attention given to a problem. You can over think things, and so can Stable Diffusion. 'Maybe this is a trick question and what the exam writer was really after was this', etc. Too many inference steps may result in additional fingers, or similar extra additions. Shown on the right is how steps effects the output of a fixed seed [4857563] a Fixed Guidance value [15] with an image size of [704-H x 512-W]

Data-set Limitations 'I want big renders' screams the endless void, well the data-set is actually 512 x 512

    

Throwing together a test machine

So I didn't grab an initial photo of this PC, but here's my test machine for synthetic diffusion, its become a bit of a 'sleeper pc'

Somewhat Basic Rectangle This isn't a remarkable PC, its a Core I5-7400, 3GHZ, but more importantly, it was immediately available. It's a retired MIT desktop PC [Aspire TC-780A] here's the build sheet [local copy]. I added in new ram, and as this only supports 2 DIMM's I added 2x 16GB  DDR4. The stock PSU is ~220W, which wouldn't work with such a large GPU. So new 650W PSU, and finally an extra leftover SATA SSD.

GPU Speaking of GPU, behold the first new GPU I've seen in ages, an RTX 3060. The goal here is to sort out how well a 'lower end' new card would work out for onboard synthetic art generation. The GPU comically barely fits inside the case, and chews up all the PCI-E spots available on the board. I got a little worried this thing would have inadequate air

Power Supply Speaking of, here's the new PSU, surprisingly good. Its quiet, the internal fan throttles down fairly low and the build quality looks quite good and only ~80USD

The weirdest Ethernet issue SU

Here is how the number of inference steps effects the final output

Lets look at this in a plot: How much does the image change step-to-step?

Concluding Remarks:

• Give

If you have questions or comments, ask below or send over an email.

Comments:

HTML Comment Box is loading comments...

Dane.Kouttron
Rensselaer Polytechnic Institute 
Electrical & Electrical Power
631.978.1650