Tristan Stevens
@tristandeep.bsky.social
Else (which I can imagine for speech) you would need a good predictive model to predict the next segment to be able to use it as initialization for the next measurement.
April 6, 2025 at 7:14 AM
Else (which I can imagine for speech) you would need a good predictive model to predict the next segment to be able to use it as initialization for the next measurement.
Interesting… not as straightforward although I’m not super familiar with speech processing. If you for instance use a spectrogram to represent the sound, and from one segment to the next these correlate, SeqDiff would be a good fit.
April 6, 2025 at 7:13 AM
Interesting… not as straightforward although I’m not super familiar with speech processing. If you for instance use a spectrogram to represent the sound, and from one segment to the next these correlate, SeqDiff would be a good fit.
Huge thanks to my amazing co-authors: @oinolan.bsky.social, Jean-Luc Robert, and Ruud van Sloun.
📄 Paper: arxiv.org/abs/2409.05399
📄 Paper: arxiv.org/abs/2409.05399
Sequential Posterior Sampling with Diffusion Models
Diffusion models have quickly risen in popularity for their ability to model complex distributions and perform effective posterior sampling. Unfortunately, the iterative nature of these generative mod...
arxiv.org
December 22, 2024 at 1:16 PM
Huge thanks to my amazing co-authors: @oinolan.bsky.social, Jean-Luc Robert, and Ruud van Sloun.
📄 Paper: arxiv.org/abs/2409.05399
📄 Paper: arxiv.org/abs/2409.05399
SeqDiff+ leverages the strong temporal correlation across frames to solve sequential inverse problems with up to 𝟮𝟱𝘅 𝘀𝗽𝗲𝗲𝗱𝘂𝗽 compared to standard diffusion models. In the example above we reconstruct partially observed ultrasound frames and compare with a vanilla DM with the same compute budget.
December 22, 2024 at 1:15 PM
SeqDiff+ leverages the strong temporal correlation across frames to solve sequential inverse problems with up to 𝟮𝟱𝘅 𝘀𝗽𝗲𝗲𝗱𝘂𝗽 compared to standard diffusion models. In the example above we reconstruct partially observed ultrasound frames and compare with a vanilla DM with the same compute budget.