Shady Abu-Hussein

Postdoctoral Researcher at the University of Cambridge (Previously: Samsung Research, Intel, IBM Research).

[Tongda Xu @xutongda]

Introducing interpolated FID: an variant of reconstruction FID that is highly correlated to diffusion generation FID.

It is well known that VAE rFID is not correlated with diffusion gFID. We propose a simple variant that achieves correlation 0.9!

https://t.co/Ml1Ig0Ocog

[Assaf Ben Kish @abk_tau]

DeciMamba, the first context extension method for Mamba, is accepted to #ICLR2025! 🎉

New revision with more long-context results:
https://t.co/MOkTS1iCW2
https://t.co/0HdaArUgIQ

Special thanks to @ItamarZimerman @ShadyAbh @nadavcohen @amirgloberson @liorwolf @RGiryes !

[Nimrod Shabtay @NimrodShabtay]

I am happy to share that LiveXiv accepted to ICLR 2025 🥳

[Nimrod Shabtay @NimrodShabtay]

Introducing LiveXiv, a new, challenging and maintainable scientific multi-modal live dataset
Paper: https://t.co/QiRrgubUom
Github: https://t.co/ttd4MYad8E
Dataset: https://t.co/4blQxQDiLz

[Raja Giryes @RGiryes]

Excited to announce that our paper on efficient diffusion models has been accepted to NeurIPS! Huge kudos to @ShadyAbh for the incredible work. Developing carefully the math and implementing it to achieve strong results took a lot of effort, and it great to see it pay off!

[Assaf Ben Kish @abk_tau]

@KyeGomezB @ItamarZimerman @ShadyAbh @nadavcohen @amirgloberson @liorwolf @RGiryes Code is now available:
https://t.co/0HdaArUgIQ

[Shady Abu-Hussein @ShadyAbh]

@fywang0126 @RGiryes In general you want to decrease the number of steps as much as you can for efficiency. However, if you want to increase it anyway, you will need to omit H in some of the diffusion steps. Which is an interesting direction that combines both worlds.

[Shady Abu-Hussein @ShadyAbh]

(1/6) 🌟 New Paper! 🌟

Is it possible to sample images from a diffusion model using less than one step?

Answer: Yes! ➡️ We present UDPM, a novel approach to image generation that outperforms existing models in efficiency and quality. 🚀📊

https://t.co/q60cCevuJJ

[Shady Abu-Hussein @ShadyAbh]

@fywang0126 @RGiryes You mean progressive GAN? it draws inspiration from it, but it is very different, we reformulate the diffusion generation scheme such that in each step you solve SR + denoise instead of just denoising.

[Shady Abu-Hussein @ShadyAbh]

(6/6)
* Check out our code and explore UDPM for your projects: https://t.co/mrMlkcJdza 💻🔗

* Read the full paper here: https://t.co/q60cCevuJJ ���🔬

#AI #MachineLearning #ImageGeneration #Research #DeepLearning #DiffusionModels #FastGeneration

[Shady Abu-Hussein @ShadyAbh]

(5/6)
* UDPM has an interpretable and interpolable latent space, making it superior to traditional DDPMs in various applications. 🌐🔍
* UDPM has smaller latent variables, hence one may modify the noise maps of each step and study its effect on the generated image:

[Assaf Ben Kish @abk_tau]

New Work! 🐍
What prevents Mamba from extrapolating to sequences that are significantly longer than those it was trained on?
Furthermore, can Mamba solve long-range NLP tasks using short-range training only?
🧵🧵🧵

[Image and Video Processing Papers @CIGX]

UDPM: Upsampling Diffusion Probabilistic Models. https://t.co/2MexBC31Ak

[Jonathan Fischoff @jfischoff]

I like the idea behind "Upsampling Diffusion Probabilistic Models" (https://t.co/xwub579qQd).

Instead using a fixed sized noised sample, use a smaller sized sample for noiser timesteps

[Shady Abu-Hussein @ShadyAbh]

@baaadas @chenlin_meng @ArashVahdat In addition to the lower number of diffusion steps, the spatial dimension of the latent variables is smaller, which allows us to better interpret the latent space.

[Shady Abu-Hussein @ShadyAbh]

@baaadas @chenlin_meng @ArashVahdat The small number of diffusion steps enables latent space interpolation