Evgenii Kashin

PhD @UniOfYork, ex-MLE @Snap, @Yandex

[Wayve @wayve_ai]

Meet GAIA-3, Wayve’s most advanced generative world model yet. 🌎

Scaling the evaluation of autonomous driving systems is one of the toughest challenges in our industry. Real-world testing matters, but it is slow, costly, and rarely captures the safety-critical events that matter most. And traditional simulation has not delivered the realism or diversity needed.

GAIA 3 changes this. It reconstructs real environments with rich fidelity and generates realistic counterfactuals that unlock safe, repeatable, scalable evaluation for modern end-to-end driving systems.

See how GAIA 3 advances Wayve toward global autonomy: https://t.co/pIk8xG1ENe

#GAIA3 #EmbodiedAI #AISafety #GenerativeAI #AutonomousVehicles

[Evgenii Kashin @digitman_]

@janusch_patas That would be perfect to have such platform for a cleaner LLM dataset

[Evgenii Kashin @digitman_]

@ericazares Out of curiosity, are you a fan of Skyrim, or were you unaware of the game?

[Evgenii Kashin @digitman_]

@jon_barron There’s an actual 4D Gaussians used for dynamic modelling https://t.co/V94RecSonk

[Evgenii Kashin @digitman_]

@janusch_patas I think a paper link is wrong, should be https://t.co/azgrze0opg

[Evgenii Kashin @digitman_]

So, in my experiments, RGB ViT beats latent ViT every time. I've tried stronger/weaker augs, bigger/smaller model sizes and image resolutions. RGB versions have ~3-5% higher accuracy than the latent ones among my experiments. You also need to use “encode” each time if using augs

[Evgenii Kashin @digitman_]

Do you think ViT trained from scratch on ImageNet would have higher accuracy in RGB or in SD's VAE latent space? In theory, the latent space should give some prior or at least works as a smart downsampling. I'm really interested in what people think. I'll share results tomorrow!

[Evgenii Kashin @digitman_]

@PDillis At the same time, the encoder has seen many more images than just 1M, if you encode/decode the images, they will look kind of the same with this compression.

Or it may act as a regularizer, preventing overfitting which ViTs are prone to

[Evgenii Kashin @digitman_]

Do you think ViT trained from scratch on ImageNet would have higher accuracy in RGB or in SD's VAE latent space? In theory, the latent space should give some prior or at least works as a smart downsampling. I'm really interested in what people think. I'll share results tomorrow!

[Evgenii Kashin @digitman_]

Generated in under 10 seconds with the vanilla SD1.5 model. I'll need to try it with more views though, WIP 2/2

[Evgenii Kashin @digitman_]

Another multi-view diffusion method? Yes! Ever since moving to the UK, I've dreamt of seeing the Stanford bunny made of baked beans 🐇 1/2

[Evgenii Kashin @digitman_]

Shout out to https://t.co/bToQ1b2Qfj for amazing pipeline. So many cool #ComfyUI workflows with #AnimateDiff

[Evgenii Kashin @digitman_]

I found cats 2024 film pipeline. AnimateDiff + IPAdapter + ControlNet

[Evgenii Kashin @digitman_]

@bilawalsidhu @fofrAI progress in the field over the last 3 years https://t.co/oafFCQ0YZE

[Evgenii Kashin @digitman_]

Accurate generation of the next frame, using several previous ones, makes it possible to predict the future or play Fortnight in a neural network. The first baseline is simply to use pix2pixHD to predict the next frame by few previous. Quality generation is still far away.

[Nikita Drobyshev @NikDrob23]

👋 Hey there! I'm Nikita Drobyshev, a dedicated Generative AI researcher with a Global Talent UK visa. I've been in London for a year, and today, I'm opening up about my pursuit for knowledge. 📚

Imagine dreaming of a UK PhD, full of ambition and credentials, only to hit a wall of unexplained refusals. Here's my reality: I've spent a year battling for my study permit (UK's "ATAS certificate") for my PhD, facing repeated refusals without any reason. 🚫🧠 Just got my second refusal, and I'm clueless why.

Guess what? I'm not alone. Many peers are in the same boat. After 3-12 months of waiting, they get "sorry" letters without an explanation. It's like solving a puzzle with missing pieces.

What's really bugging me? The lack of transparency. Imagine driving without clear road signs.🔒

Here's a one more fact: I'm Russian. Not sure if that matters, but Russian pals face permit issues since the war began. If not a coincidence, it's unfair – education should be equal, no matter your origin. ⚪️🔵⚪️

Let's set things straight: I've never been part of any big political stuff in Russia. I'm against the war in Ukraine started by the Kremlin. I left my country last year because of my beliefs 💙💛.

🚀 Let's discuss, demand clearer rules, and ensure education without barriers. Together, we can break these obstacles and prove diverse minds shape a better world.

#EducationMatters #breakingbarriers #UnityInDiversity #academia #academy #PhD #atas #Guardian #theguardian #thesun #thetimes #theindependent #ai #aicommunity #uk

[Evgenii Kashin @digitman_]

@relnox @Norod78 Wow, looks cool. But if you already have “3D conditioning” in a way of depth, you probably wouldn't benefit from PanoramaPipeline. I mean, real 3D depth should already produce results without stitching artifiacts.

[Evgenii Kashin @digitman_]

@relnox @Norod78 It probably should work with ControlNet. Have you tried MultiDiffusion (StableDiffusionPanoramaPipeline) with ControlNet?

[Evgenii Kashin @digitman_]

Try it yourself https://t.co/9yUSdDtP8R

[Evgenii Kashin @digitman_]

A few months ago, I noticed stitching artifacts while using @BlockadeLabs. I implemented a fix that seamlessly transitions from the rightmost part to the leftmost part. Although I forgot about it for a while, it's now incorporated into the @huggingface https://t.co/0UPbjZJmZU

[Evgenii Kashin @digitman_]

The original generation, without using circular_padding=True, resulted in a stitching artifact where the left and right parts didn't match seamlessly.

It just a small tweak for MultiDiffusion, for proper panoramas I'd use @BlockadeLabs (which seems to have resolved the artifact)