Global time

Global progress

Token sequence

Decoded image by level

Level 0
Level 1
Level 2

We make high-compression latent spaces competitive by enabling channel scaling.

Abstract

Fewer latent tokens promise cheaper diffusion. Yet aggressive latent compression creates a reconstruction-generation dilemma: adding channels improves reconstruction but makes generation harder. We introduce APEX, short for Asynchronous Prefix denoising for EXtreme compression, which pairs a hierarchical latent space with asynchronous denoising over channel groups. Our autoencoder orders channel prefixes from structure to detail, while one shared denoiser generates them at staggered local times without adding tokens. On ImageNet-512, APEX outperforms existing methods at every evaluated channel count under 64× and 128× compression. At matched training and guided sampling compute, it also outperforms a 256-token model at every evaluated budget, with larger gains at lower compute.

Problem

Higher spatial compression needs more channels.

Higher spatial compression makes diffusion cheaper by reducing token counts, but preserving information requires more channels.

Token count falls as channel depth grows.
Reconstruction improves, but generation collapses under matched training and sampling compute.

Hierarchical latent space

We group the latent channels into levels.

Each added level reveals progressively finer detail.

Keep through

Levels 1–2 are zeroed.

Asynchronous denoising

One global time. Different local noise.

A sampled global time maps to one noise amount for each latent level.

One shared Transformer

Sample one global time.

Result

More channels no longer mean collapse.

APEX tolerates much higher channel counts without collapsing.

Existing approaches collapse as channels grow; APEX remains stable.

Compute efficiency

Then high compression pays off.

At limited training and inference budgets, APEX makes fewer tokens outperform less-compressed models.

Guided gFID versus denoiser inference FLOPs: APEX with 64 tokens outperforms DC-AE with 256 tokens across the evaluated compute range. Dotted tails mark evaluations beyond 250 sampling steps.
APEX with 64 tokens leads DC-AE with 256 tokens across the evaluated compute range. Dotted tails mark evaluations beyond 250 sampling steps.

Examples

Staged generation: successive prefixes decoded when each level reaches t=0.
Decode each prefix as its level completes; every added level sharpens the result.
Generated prefixes decoded as each level completes.

BibTeX

@misc{dinkevich2026apexasynchronousprefixdenoising,
  title={APEX: Asynchronous Prefix Denoising for Extreme Compression},
  author={David Dinkevich and Nisan Chiprut and Yoav HaCohen and Dani Lischinski},
  year={2026},
  eprint={2610.12843},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.12843}
}