Software Engineering Courses (SECourses)•3y ago

https://www.felixsanz.dev/articles/how-stable-diffusion-works/

How Stable Diffusion works

Understand in a simple way how Stable Diffusion transforms a few words into a spectacular image.

Ffelixsanz does batch size really affects quality?

Furkan Gözükara SECourses•5/30/23, 4:13 PM

yes

Ffelixsanz but improves it? Furkan just said that a bigger batch size reduces quality!

Furkan Gözükara SECourses•5/30/23, 4:14 PM

i think it reduces for teaching a person

Ffelixsanz https://www.felixsanz.dev/articles/how-stable-diffusion-works/

Furkan Gözükara SECourses•5/30/23, 4:14 PM

thanks

Furkan Gözükara SECourses•5/30/23, 4:14 PM

i am currently focused on deep fake tutorial

Furkan Gözükara SECourses•5/30/23, 4:15 PM

FFurkan Gözükara SECourses Click to see attachment

PickledJoi•5/30/23, 5:16 PM

i don't understand. Why are you showing us a blank video?

PPickledJoi i don't understand. Why are you showing us a blank video?

Furkan Gözükara SECourses•5/30/23, 5:33 PM

this is face changed part. i will compare original and this in the youtube video

felixsanzOP•5/30/23, 5:35 PM

i was thinking this was the original lol

Ffelixsanz i was thinking this was the original lol

Furkan Gözükara SECourses•5/30/23, 5:35 PM

no you didnt watch superman :d

felixsanzOP•5/30/23, 5:36 PM

i didn't

FFurkan Gözükara SECourses this is face changed part. i will compare original and this in the youtube video

PickledJoi•5/30/23, 5:36 PM

it was a john cena joke. "you can't see him" -> blank video

PPickledJoi it was a john cena joke. "you can't see him" -> blank video

Furkan Gözükara SECourses•5/30/23, 5:45 PM

haha really :d

Furkan Gözükara SECourses•5/30/23, 5:45 PM

Furkan Gözükara SECourses•5/30/23, 5:45 PM

i am also showing face restoration

Eggo•5/30/23, 8:09 PM

Hey! Quick question. Should we use Python 3.10.6 for stable diffusion or is 3.10.11 fine?

EEggo Hey! Quick question. Should we use Python 3.10.6 for stable diffusion or is 3.10...

ashley•5/30/23, 8:11 PM

Any 3.10.x version

Aashley Any 3.10.x version

Furkan Gözükara SECourses•5/30/23, 8:50 PM

pretty much this

Furkan Gözükara SECourses•5/30/23, 8:50 PM

codeformer face fixing taking forever

Furkan Gözükara SECourses•5/30/23, 9:05 PM

felixsanzOP•5/30/23, 9:29 PM

i don't think it's a good thumbnail

felixsanzOP•5/30/23, 9:29 PM

it will be even smaller

felixsanzOP•5/30/23, 9:29 PM

it's already hard to spot the difference

felixsanzOP•5/30/23, 9:29 PM

maybe zoom on the face?

Ffelixsanz maybe zoom on the face?

Furkan Gözükara SECourses•5/30/23, 10:03 PM

how about this one

Furkan Gözükara SECourses•5/30/23, 10:03 PM

Furkan Gözükara SECourses•5/30/23, 10:03 PM

exporting the video

felixsanzOP•5/30/23, 10:09 PM

better i think, but still.... faces are too similar xD

felixsanzOP•5/30/23, 10:10 PM

but i approve!

waiting for the video

Ffelixsanz better i think, but still.... faces are too similar xD

Furkan Gözükara SECourses•5/30/23, 10:30 PM

well these guys are similar :d

Furkan Gözükara SECourses•5/30/23, 10:39 PM

uploading tutorial

Furkan Gözükara SECourses•5/30/23, 11:03 PM

video almost ready. waiting youtube to check copyright

Furkan Gözükara SECourses•5/30/23, 11:04 PM

video published. fixing subtitles and adding video chapters : https://youtu.be/OI1LEN-SgLM

felixsanzOP•5/30/23, 11:35 PM

lmao i was thinking you needed to train a face on SD or something, not just a single photo!

FFurkan Gözükara SECourses video almost ready. waiting youtube to check copyright

felixsanzOP•5/30/23, 11:36 PM

how many vram is needed?

felixsanzOP•5/30/23, 11:36 PM

only 3?

felixsanzOP•5/30/23, 11:44 PM

where did you take the model from?

felixsanzOP•5/30/23, 11:44 PM

the inswapper

felixsanzOP•5/30/23, 11:55 PM

it could be malicious!

Ffelixsanz it could be malicious!

Furkan Gözükara SECourses•5/31/23, 12:00 AM

let me find original source. on windows i couldnt make it work with gpu

Ffelixsanz how many vram is needed?

Furkan Gözükara SECourses•5/31/23, 12:01 AM

ok onnx is safe

Furkan Gözükara SECourses•5/31/23, 12:01 AM

Another mitigation approach consists in limiting what models are allowed to do, to contain and disrupt potential malware:

Some model formats are more restrictive than others, like ONNX or TFlite for instance (when avoiding custom operators).

Furkan Gözükara SECourses•5/31/23, 12:01 AM

https://github.com/Azure/counterfit/wiki/Abusing-ML-model-file-formats-to-create-malware-on-AI-systems:-A-proof-of-concept

GitHub

Abusing ML model file formats to create malware on AI systems: A pr...

a CLI that provides a generic automation layer for assessing the security of ML models - Azure/counterfit

Furkan Gözükara SECourses•5/31/23, 12:02 AM

https://discuss.pytorch.org/t/securely-serializing-loading-untrusted-pytorch-models/119744

PyTorch Forums

Securely serializing/loading untrusted pytorch models?

I am currently organizing a NeuroIPS competition, in which participants might be submitting pytorch models to our evaluation server for the leaderboard. Is there a secure way of serializing/loading untrusted pytorch models? Are there alternatively to pickle, which can be insecure? Is there a pytorch model format we can insist upon that is secu...