@justuswill.bsky.social
Joint work with my amazing collaborators Yibo Yang and @stephanmandt.bsky.social
Paper: arxiv.org/abs/2412.109...
Project Page + more details + code (available soon): justuswill.com/uqdm/
🧵4/4
Paper: arxiv.org/abs/2412.109...
Project Page + more details + code (available soon): justuswill.com/uqdm/
🧵4/4
Progressive Compression with Universally Quantized Diffusion Models
Diffusion probabilistic models have achieved mainstream success in many generative modeling tasks, from image generation to inverse problem solving. A distinct feature of these models is that they cor...
arxiv.org
April 9, 2025 at 11:55 AM
Joint work with my amazing collaborators Yibo Yang and @stephanmandt.bsky.social
Paper: arxiv.org/abs/2412.109...
Project Page + more details + code (available soon): justuswill.com/uqdm/
🧵4/4
Paper: arxiv.org/abs/2412.109...
Project Page + more details + code (available soon): justuswill.com/uqdm/
🧵4/4
Our solution? Replace Gaussian noise with uniform noise! We propose UQDM: a scalable, practical, progressive codec using universal quantization instead of Gaussian channel simulation. A single model handles a wide range of bitrates, with strong RD performance + realism. 🧵3/4
April 9, 2025 at 11:52 AM
Our solution? Replace Gaussian noise with uniform noise! We propose UQDM: a scalable, practical, progressive codec using universal quantization instead of Gaussian channel simulation. A single model handles a wide range of bitrates, with strong RD performance + realism. 🧵3/4
Diffusion models can act as progressive data compressors, allowing for intermediate reconstructions of improving quality. Prior work (Theis et al. 2022) showed it's possible in theory-but not practical, as simulating the Gaussian channel is computationally intractable. 🧵2/4
April 9, 2025 at 11:51 AM
Diffusion models can act as progressive data compressors, allowing for intermediate reconstructions of improving quality. Prior work (Theis et al. 2022) showed it's possible in theory-but not practical, as simulating the Gaussian channel is computationally intractable. 🧵2/4