Home AI Image Generation Server with LattePanda and Stable Diffusion

[@MAYERMAKES]

This is one of many easy options to get started with local Ai generation, you may score a deal on some used hardware on the local classified ads, but also keep in mind the power consumption difference of a modern Sbc compared to an older desktop pc or server. Do you have any particular applications in mind you would use image generation for?

[@andywest5773]

I really like this for images. Looking forward to better local code generation in the future. It's amazing how much AI has improved in the past year.

[@MAYERMAKES]

I think the future is in tiny models that are highly specialized snd run on minimal hardware

[@BrianMaddox]

I picked up a Tesla P40 and ended up water cooling it. There’s only so much fan noise I could stand 😂

[@MAYERMAKES]

thats a valid approach! one thing I did not do int he video for time reasosn was to use the lattepandas gpio to read temperatures and regulate fanspeed based on it to make the system more quiet.

[@BrianMaddox]

@@MAYERMAKES Yeah when I had the fan I had it set to a low speed in the BIOS then would use Coolero to spin the fan up whenever I’d do processing. Although with the fan it would get close to throttling.

[@NicksStuff]

How long to generate *one* image? Can you ask any resolution?

[@MAYERMAKES]

on resolution: the models running on this usually expect 512x512 or anything close to it for portrait or landscape oreintation), you can later upsacale the selected good images. The generation time is dependent on the exact settings and models used. every sampler/model combo needs a different amount of iterations to yield decent results, so it can range form a few seconds for a single image to multiple minutes. A common realistic model with the sampler "restart" and 20 iterations on 512x512 nieds about 30sec to 1 minute per image. generating multiple images wiht the same paramters at once is usually faster then generating them one by one, but needs more Vram available. with the 8Gb of Vram this system has 4 images per batch works fine with decent times.

[@NicksStuff]

@@MAYERMAKES 20 iterations on 512*512 images? Is that 30 sec to 1 min per iteration? Or is 20 iterations a parameter that gives you your image within a minute? That you then need to upscale

[@MAYERMAKES]

Thats for the 20 iterstions, resulting in one image0​@@NicksStuff

[@NicksStuff]

@@MAYERMAKES Thanks!

[@RodgerE2472]

Looks like fun, wish I had the cash. 😊 Great video!

[@MAYERMAKES]

The overall cost was surprisingly low, since a lot of parts were used before. You should be able to build this for under 500€ total.

[@whatthefunction9140]

I just got 3 of these old cards on Amazon for 49 each

[@MAYERMAKES]

Scooore😮

[@linuxguy1199]

On a budget lol, for the price of that LattePanda you could've got a used HP Proliant DL360 off eBay, and had a few bucks to spare! I got one at a hamfest for $35 that'd be prefect for this, plus I already have a box of 20 Tesla P100s I got for free from a friend of mine.

[@MAYERMAKES]

there is always the possibility of finding a better deal, but an Sbc also consumes a lot less power than an old desktop and when operating it continously that adds up. Also showing an option that anyone can easily obtain (including warranty etc) makes a bit more sense in a video than just showing off a great deal form the local classified ads.

[@luisdaniel3363]

let's save to computer to spent 3x on a GPU. Even lattepand is not cheap and with more few money can buy a new cheap computer and do not need extra tools. If sum the money you spent with converters will be almost the same as a cheap computer where can attach CPU within the case.

[@MAYERMAKES]

The whole thing was less than 500€ , if you wanted you could mount that in a tiny case. There are plenty ways to skin a cat, but the polnt ks to show how you can add a gpu to an sbc quite easily. And on top the lattpeanda uses very little power in comparison to a normal PC.

[@qeqsiquemechanical9041]

AI craze is starting to get old. Ew.

[@MAYERMAKES]

Just now the models are getting good and small enough to run locally. the craze is for those who arent able tio utilize it meanwhile people building their own systems get ahead and train bespoke models locally.

[@JJSideshowBob]

Tough times ahead of you.

[@gothixxx12]

So leave and unsubscribe.

[@qeqsiquemechanical9041]

@@gothixxx12 i don't need any advises from ai techbro, stick 'em up you know where