Lightricks' APEX cuts diffusion to 16 tokens, beats baselines

Latent-diffusion image models in the SD-VAE lineage operate on 1024-token grids. A new paper from Lightricks and Hebrew University of Jerusalem reports matching that quality with 16. The authors, David Dinkevich, Nisan Chiprut, Yoav HaCohen and Dani Lischinski, call the system APEX, and in the arxiv preprint they frame the core difficulty as "a reconstruction-generation dilemma: adding channels improves reconstruction but makes generation harder."

Their fix pairs a hierarchical autoencoder with a single denoiser that handles channel groups at staggered noise levels. "Our autoencoder orders channel prefixes from structure to detail, while one shared denoiser generates them at staggered local times without adding tokens," the abstract states.

The headline result sits at 128× spatial compression, where a 512×512 image collapses to a 4×4 grid of 16 tokens. "At f128c512, APEX reaches 8.98 guided gFID from 16 tokens with SiT-XL, while every compared method remains in the double digits," the paper reports. At the less aggressive 64× setting, APEX's f64c128 variant records 2.85 gFID against DC-AE's 3.22; at 128×, DC-AE 1.5 lands at 11.86 and the original DC-AE at 14.44. The authors say APEX matches a published 1024-token SD-VAE result using "15% of its denoiser-training FLOPs."

The abstract names only ImageNet-512 as its evaluation benchmark, and the baseline comparison numbers come from the authors themselves.