Lightricks' APEX pairs hierarchical latents with async denoising for extreme 64-128x diffusion compression

Lightricks · hf · 2026-10-14

Lightricks and Hebrew University introduce APEX (Asynchronous Prefix denoising for EXtreme compression), tackling the reconstruction-generation dilemma of aggressive latent compression.

• The autoencoder orders channel prefixes from structure to detail; one shared denoiser generates channel groups at staggered local times without adding tokens
• Naively combining channel hierarchy with async denoising hurts; success comes from matching what each level represents, when it denoises, and training allocation
• On ImageNet-512, APEX beats prior methods at every channel count under 64x and 128x compression, and outperforms a 256-token model at every compute budget, with larger gains at lower compute

Code is open-sourced.

Original post →

More from Multimodal

• Video models keep sneaking in music; Mirelo SFX 2.0 removes just the music — SimplyAnnisa · 2026-10-14
• Mirelo SFX 2.0 launches on fal: swap video sound effects while keeping dialogue — SimplyAnnisa · 2026-10-14
• Omni-LLMs Know More Than They Say: NAVER's OmniCatalyst Adds 7.29% via Self-Generated Text — naver-ai · 2026-10-14
• Meow: AI-Generated Cat Art Turns Whiskers Into Streaks of Light — gizakdag · 2026-10-14
• Breezeblue-TTS-2 claims best local TTS crown, beating Moss TTS 1.5 but non-commercial — 9r4n4y · 2026-10-14
• Stunning AI-Generated Sci-Fi Shots of Stealth Jets at the Edge of Space — egeberkina · 2026-10-14

Multimodal channel →