Papers of the day   All papers

Generating Diverse High-Fidelity Images with VQ-VAE-2


Aäron van den Oord: VQVAE-2 finally out! Powerful autoregressive models in a hierarchical compressed latent space. No modes were collapsed in the creation of these samples ;) Arixv: With @catamorphist and @vinyals More samples and details 👇 [thread]

12 replies, 906 likes

Oriol Vinyals: Surprising how simple ideas can yield such a good generative model! -Mean Squared Error loss on pixels -Non-autoregressive image decoder -Discrete latents w/ straight through estimator w/ @catamorphist & @avdnoord VQ-VAE-2: Code:

7 replies, 662 likes

Oriol Vinyals: Great post by Prof. David McAllester on why discrete representations matter, based on our findings in VQ-VAE2. "Vector quantization seems to be a minimal-bias way for symbols to enter into deep models."

4 replies, 406 likes

Ben Poole: Big hierarchical VQ-VAEs with autoregressive priors do amazing things. Awesome work from @catamorphist @avdnoord @OriolVinyalsML:

2 replies, 324 likes

roadrunner01: Generating Diverse High-Fidelity Images with VQ-VAE-2 pdf: abs:

2 replies, 127 likes

Gene Kogan: For most of the creatives/non-scientists out there, this may seem like just another BigGAN/StyleGAN, but this has important advantages: It's likelihood-based (can be evaluated formally), samples much faster, and should be superior in generator diversity. Really good stuff

3 replies, 114 likes

Xander Steenbrugge: Generative Modelling space on fire! After Google's #BigGan and Nvidia's #StyleGAN we now finally have autoencoder based models that generate samples of equal/better? quality! The sample diversity is especially striking given that mode collapse has always been an issue for GANs.

0 replies, 51 likes

Kyle McDonald: VAE-style networks have surpassed the quality of BigGAN and StyleGAN. i always knew they had it in them 🎉

1 replies, 43 likes

Max Jaderberg: Insanely good samples from the latest incarnation of the VQVAE generative model

1 replies, 40 likes

François Fleuret: Beside the quantitative evidences that they are more robust to mode collapse than GANs, their roots in "classical" density estimation make VAE more promising as a generic tool. We have "good enough classifiers" since 2015, maybe are we also good for density models...

0 replies, 27 likes

Danilo J. Rezende: Great results on generative modelling from @catamorphist, @avdnoord and @OriolVinyalsML !

0 replies, 21 likes

Simon Kornblith: @carlesgelada @timnitGebru @ylecun Here's my favorite example of this, from The left are samples from VQ-VAE2; the right are from BigGAN; both are trained on ImageNet. Isn't it kind of obvious what's going to happen if these algorithms are a trained on a face dataset?

3 replies, 18 likes

Daisuke Okanohara: VQ-VAE-2 improves VQ-VAE by using1) hierarchical latent variables 2) a prior distribution that matches the marginal posterior using an auto-regressive model with self-attention; achieving diverse and high-fidelity image generation.

0 replies, 15 likes

d00d: VAE based image generation with quality comparable to GAN generated images, but more variety and faster sampling...

1 replies, 12 likes

Alex Nichol: The VQ-VAE-2 paper is hilariously vague. E.g. "It consists of a few residual blocks followed by a number of strided transposed convolutions". (Paper:

1 replies, 6 likes

René Schulte: Impressive new step for generated images. The below photos are all synthesized by an AI 👌 Instead of a GAN they use a Vector Quantized Variational AutoEncoder (VQ-VAE) which makes it easier to handle and much faster. 🚀 #AI #DeepLearning #ML #DNN

0 replies, 6 likes

Kaixhin: I love this mix between very general (autoregressively-decoded discrete sequences), general (hierarchical structure) and specific (local spatial structure) priors to model complex distributions in the real world 🌏

0 replies, 6 likes

Alex J. Champandard: 2/ At this stage, we know it's possible to generate HD images with many/most techniques. NVIDIA built StyleGAN, OpenAI developed GLOW, DeepMind created VQVAE, etc. Everyone has their favorite! 🐕 .

1 replies, 4 likes

Alex J. Champandard: 6/ The idea of working in a smaller and coarser space is not new. It's what made GANs scale to 1024x1024 in the first place (progressive growing) and it's the idea that helped VQVAE catch up.

1 replies, 3 likes

Kyle Kastner: @selimonder This paper (along with a few others recently such as , are exploiting the multi-scale structure inherent in audio and images. That kind of structure is much harder to get *easily* in language - dependency which may not be local

1 replies, 2 likes

Seth Stafford: THIS: "The . shift from symbolic logic . to distributed vector representations is . viewed as [a] cornerstone of . deep learning . I . believe . logical symbolic reasoning is necessary for AGI. Vector quantization seems . a minimal-bias way for symbols to enter . deep models."

0 replies, 1 likes

JFPuget Wash Hands Social Distancing Wear Mask: A clear example of two different ML algorithms. The one on the left has much more diverse outputs (and better quality ones) Not sure though it is less biased.

0 replies, 1 likes


Found on Jun 04 2019 at

PDF content of a computer science paper: Generating Diverse High-Fidelity Images with VQ-VAE-2