hello!
i am back after an long time and i want to tell about my home lab!
well first of all i have 2 aruba AP’s i am currently using them with POE+
and i have 200 switch ports all aruba and HPE and enterwan
and my router is an raspberry pi 5 with an pro/1000CT NIC shoved into it
and my server is way bigger now
it has 2x E5 2680 V4 with 56 theards
64 gb of ECC ram 8x8 B die samsung ram
20TB of storage powered by ZFS
and some SAS equidement
and i got an mointor from the trash its 1080p@144HZ
i replaced my uap ac pro with aruba!
and by the way these exploits i found
i found out how to plat minecarft since my school left an CIP file thatgave sepfick path’s infinite power
and bypass WDAC
and on marktplaats i found out if you search the API they leak the sellers adderes (bad right?)
and also i made an new AI arch all by my self
look
“””
veiter
AI image generatoin thingy this is meant to be faster better and fight nightshade meant to be only used on anime
what is nightshade?
stupid tool that artists use XXXX the AI and wil cook an ai
=======================
why artists use it?
they want no ai traning on them art
+++========================
how to solve this probelm
we replce CLIP with an "decompiler" clasfiyer
this clasifyer has an Vit AND 3 MoE , the ViT is trained on milions of images on ONLY booru nofing else
and e1 says if its s q e or g e2 says if its ai slop or not ( we wil scarpe tons of ai slop sites via the RE'd api or just the offcial one and then tons of human from danbooru and then let it classify if human or not ) e3 says the booru tags like it wil look at an image and flash hatsune_miku slot at 0.97 and chibi at 0.96 and jeffrey_epstein at 0.00000000000000000001 the output wil be hatsune_miku, chibi this is due to it being trained on rule34 danbooru gelbooru e621 realbooru etc
and almost all the booru's have an API and they have milions of images if some don't you can RE it or use gallery-dl if you get the imges you gotta dedupe a LOT of them and then also do image coruption checker and now u need to remove any images that have stupid tags or move them like ai_generated (feed it to e2)
now VAE is next we replace it bytedance titok (not tiktok) that is 410X faster and ueses an ViT instead of an VAE (wil also be trained purley an anime) ( https://github.com/bytedance/1d-tokenizer )
and the UNET replaces with an DIT
(the loss func i am still thinking about perceptual loss )
but the loss fungsoin is inportned the ai wil try to geuse out of the noise but an DINOV2 and LPIPS finetuned purley on anime it wil comapre the 2 and give an loss fungsoin
================== bootstrap
we now scarpe major sites liek pixi etc and with that art we use the decompiler to filter dlop and tag it and rate if its porn or safe (e1) and then we retrain (we only get images that are +0.9)
now we are god
============ why anime?
becuz booru tags are god mode
======================== why do we need this?
improve performance and solve the nightshade probelm
“””
Moderator edit: het is niet toegestaan om te schelden op de community. Dit is ook meteen je laatste waarschuwing.
