APEX launches asynchronous prefix denoising for extreme compression
| Source: HF Papers | Original article
Researchers unveil APEX, an asynchronous prefix denoising method that enables extreme latent compression while balancing reconstruction quality and diffusion generation efficiency.
A new paper titled **APEX: Asynchronous Prefix Denoising for Extreme Compression** proposes a radical shift in how diffusion models handle latent representations. The authors combine a hierarchical latent space with an asynchronous denoising schedule that processes channel groups in stages. By allowing a single token to receive multiple diffusion timesteps simultaneously, the method first generates a coarse image in a simplified space and then refines it with finer details.
The technique delivers “extreme compression” – achieving 64‑fold and 128‑fold reductions on ImageNet‑512 while still outperforming previous approaches. Crucially, APEX breaks the long‑standing reconstruction‑generation dilemma: adding channels traditionally improves reconstruction quality but makes the diffusion process harder. Through level‑aware weighting of a shared denoiser, the model attains comparable training and inference compute to prior work but with **four times fewer tokens**, as highlighted by the authors’ tweet: “We give the same token multiple diffusion timesteps at once. APEX delivers better generation at the same compute with 4× fewer tokens.”
Why it matters is twofold. First, token count is a primary driver of diffusion‑model cost; cutting it dramatically lowers inference expenses and energy use. Second, the hierarchical‑asynchronous scheme shows that aggressive latent compression need not sacrifice fidelity, a theme echoed in our earlier coverage of sub‑1‑bit LLM compression (Oct 9) and token‑compression tricks for coding agents (Oct 1). APEX therefore pushes the frontier of affordable, high‑quality generative AI.
Looking ahead, the research team has opened a project page and released code, inviting the community to test the approach on other datasets and modalities. Watch for integration of APEX‑style denoising in upcoming image‑generation APIs, potential extensions to video synthesis (building on work such as GRACE), and benchmark updates that could reshape cost models for large‑scale diffusion deployments.
Sources
Back to AIPULSEN