>>188864 #
qwen image-edit but it's still censored, true open-source solutions are always downstream as they take a lot of MONEY to train to desirable capability, which is only done by individuals who aren't already utilizing their skillset to work for the very companies that produce the aformentioned

>>188880 #
makes sense

You need to know how to wrangle nano banana a bit but it's honestly amazing if you also use the general image gen capabilities

For example, nano banana inpainting will happily spit out the same image if your instructions weren't transformative enough or if they're not allowed for whatever reason, despite Google's official reference, you shouldn't refer to 'provided images' or 'this image' for your first image, which is what it uses as primary target, doing it for second and further images is fine

'create another image like this' or 'change this image' has it write off your edits far too quickly, if you treat it like chatgpt and just let it infer what you want from the images provided, it's much more versatile

What's up guys first of heil Hitler and then secondly nano banana is bad ass! Thank you for image reference inference yahooo

milady

chat-gpt-5:high is the most capable LLM change my mind

bro is using an LLM to analyze tweets and send DMs

does anyone know how to inverse a ddpm

pee pee

milady

milady

milady

poo poo

nigga AI board
I like the darkness

[YouTube] ???

penis

mlady

h

i

omg hello /ai/

haven't looked myself, but Im curious about research into model degradation
like there's quantization which is a controlled way of shrinking a model size, but what about taking a weights file and removing bits and testing output
sort of a 'how much sawdust can you put in a rice treat before noticing' type experiment

>>272644 #
https://arxiv.org/abs/2505.13500

>their resilience under perturbation remains poorly understood

Spent a few minutes looking on arxiv and found this one which is what I was thinking about, should be easy to mess around with with local llms

Im getting more ideas now, what if you had an unperturbed LLM testing perturbed models in a loop where you're testing for novel outputs and save perturbations and using genetic algos mate interesting ones together to produce some real interesting behavior

oh no my big uh ohmm uhh

hi

hi

>>188923 #

test

test

hello

beep boop

shimmy shimmy ye

Hello

>>308297 #
>>308297 #

boop beep

gmeow

fart ai

B

AI

gm

milady

8

yodie gang

wfer

asdfasdfasdfasdfasfasdfasdf

is there a way to create private chats on here since somachat is permanently down

Z

>>313664 #

miFUCKINGlady

milady

milady

milady

milady

milady

milady

Does anyone have a decent guide for local-hosted video generation? I've been using Grok Imagine and I'm pleased with the results but I'm curious what other options look like

milady

>>378548 #
ComfyUI?

>>364642 #

hi, does anyone have tips for scraping tweets from profiles with grok? as theres no public api seems like easiest way to go about it is via grok