Published on:
8 May 2024
Primary Category:
Computer Vision and Pattern Recognition
Paper Authors:
Jonas Kohler,
Albert Pumarola,
Edgar Schönfeld,
Artsiom Sanakoyeu,
Roshan Sumbaly,
Peter Vajda,
Ali Thabet
Proposes distillation framework tailored for extremely low-step diffusion model inference
Introduces Backward Distillation concept to calibrate student on its own backward path
Presents Shifted Reconstruction Loss to transfer hierarchical knowledge
Applies Noise Correction to address singularities in early sampling
Achieves teacher-level performance in just 3 steps
Accelerating diffusion models with distillation for fast high-quality image generation
This paper proposes a distillation framework to accelerate diffusion models, enabling high-quality and diverse image generation using only 1-3 sampling steps. Key innovations include Backward Distillation to reduce train-test discrepancy, Shifted Reconstruction Loss to transfer both structure and detail knowledge, and Noise Correction to enhance initial sample quality.
Accelerating diffusion models for fast image generation
Fast super-resolution with stable diffusion
Real-time high-fidelity image generation
Fast text-to-image generation in one step
Efficient synthetic training data generation through expert-student alignment
Efficient conditional diffusion model for image super-resolution
No comments yet, be the first to start the conversation...
Sign up to comment on this paper