Introducing "Spectrally-Guided Diffusion Noise Schedules", accepted to #ICML2026! With @kiamada.
We propose noise schedules that follow the power spectrum of each image, improving pixel diffusion quality and requiring fewer denoising steps.
At test time, we conditionally sample
We'll present "Spectral Image Tokenizer" at #ICCV2025 later today, afternoon session.
We tokenize the image spectrum, train an autoregressive transformer for coarse-to-fine generation, and show applications to image generation, upsampling and editing.
w/ @kiamada@msuhail153
Our new paper, "Spectral Image Tokenizer", is on arXiv! We train a tokenizer on DWT coefficients
that enables autoregressive coarse-to-fine image generation, w/ applications to multiscale text-to-image, and text-guided editing.
w/ @kiamada, @msuhail153arxiv.org/abs/2412.09607
Our new paper, "Spectral Image Tokenizer", is on arXiv! We train a tokenizer on DWT coefficients
that enables autoregressive coarse-to-fine image generation, w/ applications to multiscale text-to-image, and text-guided editing.
w/ @kiamada, @msuhail153arxiv.org/abs/2412.09607
Applying computer vision models designed for planar images to data projected on spherical surfaces is challenging. Here we present an open-source library in JAX to solve the challenges of rotation and regular sampling for state-of-the-art performance → goo.gle/46z3vD7
