haru

@kurumuz you dont understand i just love suffering github.com/harubaru/vkwai…

lets make rendering a triangle to the screen super hard just to have fun

@linaqruf_ additionally, the "we need the model!!" attitude is a large portion of what drove me away from open source AI for the large part among other issues. i could *really* care less nowadays about what people do or say nowadays since i don't have involvement in projects like that.

@linaqruf_ yeah i know good models are being trained on it, i personally just don't like sdxl. people can do whatever they want that's not really my business.

@main_horse @linaqruf_

@haruu1367 @linaqruf_ really now

Telling my kids this was sam altman

@JesseCohenInv okay this is peak silliness

Wow. This is big if true. $NVDA There's a mind-boggling theory that's gaining momentum on X/Twitter that Nvidia's blowout performance this year has been fueled by fake GPU sales to fake non-existent clients. Some have speculated that an AI tech company called #Coreweave has made orders from Nvidia amounting to billions of dollars. There's only one caveat though - Coreweave might not even be a real company!! 🚨🚨 Instead, Coreweave is actually a shell company created and backed by Nvidia and Blackrock, which is Nvidia's largest shareholder. Some have started to make parallels and comparisons to the recent FTX ponzi scheme debacle, which resulted in the crypto company's collapse. A scam like this can take down the entire market with it. Remember, there's no smoke without a fire. $NVDA Good job to all those involved in bringing these developments to light @SamanthaLaDuc @JG_Nuke

@mkualquiera godspeed 🫡

@haruu1367 Trust me bro, I know

i think the right hyperparameters were found after a lot of runs.

@kaj718 tensormancer should be as valid as machine learning engineer

I don't care if "technomancer" is cringe I think it sounds cool and awesome and it should be just as valid as "software engineer"

was #SDXL trained to condition on *both*: - [left] empty-string uncond - [right] all-zeros uncond ? they resolve to similar images, but all-zeros looks more in-distribution. there isn't always a clear winner, but over a few cases their output tended to be similar.

@Birchlabs @s_alt_acc @kohya_tech i wish i knew where to begin finetuning the refiner lol

@haruu1367 @s_alt_acc @kohya_tech looking forward to finetuned text encoder! I think CLIP ViT-L and ViT-bigG don't have good embeddings of the characters I've tried. are you going to train a refiner? #SDXL's current refiner already adds some nice high frequency details: left = base right = base+refiner ensemble

I've converted waifu-diffusion-xl for use with diffusers. thanks @haruu1367 @s_alt_acc, closertodeath, Kudo and anyone else who helped with training. コンバータスクリプトをつくのが@kohya_techに感謝している。 ensemble with #SDXL refiner still seems to work. huggingface.co/Birchlabs/waif…

@araminta_k We will publish our process for training it once we release the model.

@haruu1367 Oooh, I'm so curious to know how the process is going! Going to have to finetune a lot myself :)

WDXL. 5% trained.

@kurumuz true, i'm getting flashbacks to when google finetuned an image gen model on one image and somehow turned it into an 18 image paper

bro why is everything a paper now, just normalize blog posts

anyone wanna help me with RLHF experiments for instruct-tuned models? i'm not too sure what i am doing but so far i think the model is working well, any help would be much appreciated!

@DavidSHolz nope, it generated the descriptions from just the images alone. no prompt at all

@haruu1367 did you ask it a particular prompt to get these descriptions?

GPT-4 has some really interesting vision capabilities especially for describing art.