What does it mean for an image, video, or text to be 𝑟𝑒𝑎𝑙𝑖𝑠𝑡𝑖𝑐?. Despite how far we've come in 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑛𝑔 realistic data, 𝑞𝑢𝑎𝑛𝑡𝑖𝑓𝑦𝑖𝑛𝑔 realism is still a poorly understood problem. I've shared my thoughts on how to correctly quantify realism here:.

RT @sedielem: #ICML2024 diffusion circle 📢: let's meet at 3PM on Thursday at registration, near the Vienna tourism desk. We'll find a spot….

RT @elluba: The @CVPR AI Art Gallery is now live 🤖🎨. Featuring 115+ artworks using or about computer vision😎. View them here: https://t.co/….

I am at CVPR to give a talk at the AI for Streaming workshop ( and to present our paper on C3. Let me know if you’re in Seattle and would like to catch up. #CVPR2024.

RT @mhutter42: Don't miss this ICML Spotlight Talk by one of my colleagues, and/or even better read his very nice exposition illustrating w….

RT @emidup: We build neural codecs from a *single* image or video, achieving compression performance close to SOTA models trained on large….

RT @Official_CLIC: CLIC is back! Details are now online @ The validation phase starts in 3 weeks. Get ready!🏎️.

RT @karen_ullrich: 🤩 Tomorrow the #ICML2023 workshops shall begin.🎙️ Join me at 9 AM I will discuss absolutely all there is to know about #….

RT @george_toderici: CLIC (Challenge on Learned Image Compression) is not dead. Stay tuned for some exciting news.

w/ @emiel_hoogeboom @etagust @mentzer_f Luca Versari @george_toderici.

Arbitrary resolutions are handled by generating images patch by patch. This makes it easy to use network architectures that don't scale or generalize well to arbitrary resolutions (e.g., attention layers).

On the other hand, we show that a simple two-stage diffusion approach can be made to work very well with minor tweaks (e.g., to the noise schedule).

As others have found, beating well-tuned adversarial approaches to compression (e.g., HiFiC, PO-ELIC) with diffusion is challenging. Repurposing text-to-image models (e.g., Stable Diffusion, Imagen) does not work well as these have not been designed to reproduce fine details.

New paper on high-resolution image compression with diffusion models, achieving SoTA results in terms of FID:.

RT @_akhaliq: High-Fidelity Image Compression with Score-based Generative Models. Despite the tremendous success of diffusion generative mo….

For a distribution which is heavy-tailed enough, you should expect to wait forever.

This can be viewed as an extreme version of the waiting paradox. Even when the average distance between buses is the same, the waiting time can differ significantly depending on the 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 over distances.

Worse, the 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 runtime can be infinite when averaged over target distributions.

Surprisingly, although the coding cost of "greedy" rejection sampling (Harsha et al., 2008) tends to be much lower than rejection sampling, the runtime is identical, namely exponential in the ∞-divergence between the proposal and target distribution.

New paper with Greg Flamich on how to communicate samples efficiently (or 𝑐ℎ𝑎𝑛𝑛𝑒𝑙 𝑠𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛). In addition to proposing a fast new scheme, working on this paper clarified a few things for me about the runtime of existing schemes.