Sabitlenmiş Tweet
Max
2.4K posts
Max
@max_pe2002
🇪🇺eu/acc, compressing reality (pied piper)
earth Katılım Ağustos 2020
661 Takip Edilen236 Takipçiler
@SwayStar123 @NicholasBardy @JiaweiYang118 also doesnt work for conditional from scratch training as it is currently. would have to seperate FID over classes
English
@NicholasBardy @JiaweiYang118 Not sure that this is stable, they showed some samples from overtraining, it collapses/reward hacks. So i dont think this would work for pretraining. Correct me if im wrong
English
Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space.
Now it is 0.75, and can be even lower.
Many wonder how.
I thought it might end as a small FID prank: simple and deliberate.
It started with one question: can FID be optimized directly, and what does it reveal?
Introducing FD-loss.
English
@alex_peys yeah this paper probably works as a straight path from noise to sample. but having to run like 3 encoders might be slower than just training with mse
English
a good question asked and answered: why are you training with one metric (e.g. mse on flow) and evaluating on another one (FID)?
Jiawei Yang@JiaweiYang118
Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space. Now it is 0.75, and can be even lower. Many wonder how. I thought it might end as a small FID prank: simple and deliberate. It started with one question: can FID be optimized directly, and what does it reveal? Introducing FD-loss.
English
@sama @lukaszkaiser claude code is better but if you actually want to get done with more than 2 back and forths codex is better
English
you know what
all of these "which is better" polls are silly
use codex or claude code, whatever works best for you
i am grateful we live in a time with such amazing tools, and grateful there is a choice
English
@JiaweiYang118 @ChongZitaZhang very very good work! congrats! reminds me a bit off the Drift paper but better
English
was debugging my x0 trainer using 1 image with lots of high frequency in it and got this interesting pattern.
it seems to converge on high freq first before everything else.
seems like x0 is doing frequency weighting implicitly based on your data distribution.
English
I can't believe we were right
Claude was dumbified on March 4, just when we noticed!
@levelsio@levelsio
Claude Code with Opus 4.6 was so dumb today I finally had to write my own code again A sad state of affairs 🥹
English
@madebyollin @gabeeegoooh @SwayStar123 hmm seems like a GAN artifact then if its always the same position
English
@max_pe2002 @gabeeegoooh @SwayStar123 Hmm, have all of your artifacty images been from multi-image chats? It looks like there's a strong bias towards texture/feature copying across images within a chat thread:
x.com/madebyollin/st…
Ollin Boer Bohan@madebyollin
@JiaweiYang118 Some of the ChatGPT Images 2 weirdness is texture leakage from context images (see this thread reddit.com/r/ChatGPT/comm…) You can visualize this by asking for three unrelated images in the same chat (note how the flower texture persists in the circled area).
English
Turns out I invented gpt image 2 before openai did
sway@SwayStar123
If you distill a model with LADD at low resolution, then use it at high resolution, the model is heavily biased towards highly detailed images!
English
English
@max_pe2002 This reddit thread has more examples of the weird mid-frequency artifacts, seems like a widespread issue... reddit.com/r/StableDiffus…
cc @gabeeegoooh for awareness
English
@madebyollin yes but i also see like a repeating thing at the bottomo of the first image
English
@max_pe2002 Yeah looks like LDM decoder, maybe with some sort of scheduler issue (not enough mid-frequency noise removed?).
English
It's very obvious they specifically trained it on millions of software interface screenshots.
Now ask yourself why they would do that
Kate Deyneka@katedeyneka
GPT-Image-2 is out! the model is insanely good at rendering text and generating all the tiny details in complex software interfaces - I'm super impressed prompt: “Generate a realistic desktop screenshot of Adobe Premiere”
English
@industriaalist omg finally someone who understands. with limited data and unlimited compute diffusion models are better but per data they are worse.
English
@SawyerMerritt yeah still not upgrading my tesla until hw 5 is in cars after what they did to hw3 users
English
KI Modelle werden immer stärker. Die Unterschiede sind so groß dass man sie beim Wechsel auf bessere Versionen sofort erkennt. Niemand will daher ein Modell der 2. Liga. Es scheint mir katastrophal, dass wir mittlerweile hier komplett abgeschlagen sind. Eine enorme Abhängigkeit
Antonin Bergeaud@a_bergeaud
Le AI index report de Stanford (@erikbryn et al.) est sorti aujourd'hui et le premier graphique est assez douloureux hai.stanford.edu/ai-index
Deutsch
@neural_avb should play boss music as soon as he says my name is jürgen schmidhuber
English
My favourite piece of Schmidhuber lore is when he challenged Ian Goodfellow during a NIPS presentation on GANs
Live in public
Deep Learning drama peaked here. You have seen nothing like this.
Yuntian Deng@yuntiandeng
Glad to see followups to neural-os.com, but disappointed that neither the blog (with 34 refs) nor the code repo acknowledged NeuralOS, even tho the released data code appears to build directly on top of ours. That omission is hard to understand given our shared vision.
English