To anyone who still considers PSNR to be a useful metric to assess image quality, please take a close look at this:
https://t.co/LCf6jHN1sP
According to PSNR, those four images have the same quality. According to my eyes, they don't.
That's the beauty of it bro, we study the source material, page after page, book after book... They won't even know we're actually larping...
And when the questions roll in, we use that knowledge to our advantage and we can keep posing like nothing even happened...
@HissHitbox@skxllflower@HunterBiden@YouAreWrongOk i have no idea what llm you've been using, but probably every model released by a lab in the past 3 years doesn't use hyphens like that without a hyper-specific system prompt to make mistakes as a human would.
@rnaferreira@NexEcosystem While it's true that by itself it proves nothing, the weights being rather exact element-wise as proven by a simple linear interpolation is surprising. Apparently they uploaded the 'base' model instead of their finetuned version though.
@donttagme6969@deanwball having knowledge of something isn't "the tools to build"
the knowledge of how to build a bio-weapon isn't the barrier, it's the investment, production, and deployment of one.
For 7k steps into a run of a <1M parameter SISR arch, it's actually seemingly usable on vgg19!
Gonna try it out with a deeper stage of DINOv3, see if that does anything special 👀
(lr 4x nearest versus sisr 4x)
ik its not too useful without relative comparisons, sry for now
🚨 New paper: Introducing MIND (Monge Inception Distance)
Everyone agrees that FID is broken, requires too many samples, slowing down evals.
MIND requires 10x fewer samples, is more robust, faster to compute.
Our new drop-in replacement for evaluating generative models. 🧵👇
🚨 New Paper 🚨
ScheduleFree+: Scaling Learning-Rate-Free & Schedule-Free Learning to Large Language Models
A few modifications to Schedule-Free Learning make it completely LR tuning free, and allow it to greatly outperform schedules for long duration training!
https://t.co/LzjIIsOlG8
Kids are smoking AI diffusion models.
This comes after a typo in a tweet from the W.H.O. ("I've been legit perma-tripping since I smoked Klein 9B")
Footage emerged of a BFL employee stuck in the PCA of block 11 in a LDM
And to make matters worse
Emad is coming back to StabilityAI