Consistency Distillation for EDM (ImageNet-64, conditional) — In Progress

This repository is now focused on implementing Multistep Consistency Distillation (MCD) in the EDM framework for ImageNet-64 with class-conditioning. The legacy EDM README has been moved to docs (docs/README-legacy-edm.md)

Status

• Target: S-step deterministic student (S ∈ {4, 8, 16}) distilled from a frozen EDM teacher using Heun integration over a denser Karras sigma grid, in EDM parameterization.

What we’re building

• Consistency distillation in EDM space with:
  • Deterministic Heun teacher hops
  • Student single-eval Euler/DDIM forward
  • Segmenting the teacher grid and pushing to student boundaries
  • invDDIM targets and EDM-weighted Huber/L2 loss
  • Full DDP support, augmentation compatibility, and annealing of teacher edges

Citations

• Multistep Consistency Models:
  • Heek, Jonathan, Emiel Hoogeboom, and Tim Salimans. “Multistep consistency models.” arXiv preprint arXiv:2403.06807 (2024).
• EDM:
  • Karras, Tero, Miika Aittala, Timo Aila, and Samuli Laine. “Elucidating the Design Space of Diffusion-Based Generative Models.” NeurIPS 2022.

Datasets and evaluation

• ImageNet-64 preprocessing and FID computation remain supported; see dataset_tool.py and fid.py.
• During development, we will periodically evaluate S-step deterministic sampling and report FID against imagenet-64x64.npz.

License

This repo inherits the original EDM license terms. See LICENSE.txt.