superseded by latent diffusion, which denoises in a compressed VAE latent instead of at full resolution; pixel-space diffusion survives inside cascaded super-resolution and relay upsampling stages
Learn more →
Why study this historical topic?
superseded by latent diffusion, which denoises in a compressed VAE latent instead of at full resolution; pixel-space diffusion survives inside cascaded super-resolution and relay upsampling stages
Diffusion Trained Directly in Pixel Space
denoising at full resolution is the obvious first design and its compute cost is exactly what latent diffusion exists to fix
Predict Geometry Not Velocity Using PointDiT Pixel Space Diffusion in Computer Vision
The core theoretical principle is "X-prediction" in pixel-space diffusion models for monocular depth estimation, which posits that direct prediction of clean dense 3D point maps yields superior geome…