Recurrent self-conditioning: framing recent diffusion methods as depth decoupled from time
alec_helbling · x · 2026-10-14
In a discussion on diffusion language models, Alec Helbling argues recent approaches can be understood as "recurrent self-conditioning": the model gains an extra depth component decoupled from diffusion time.
In conventional self-conditioned sampling each pass advances diffusion time; here, clean predictions are refined for K steps without advancing it. Replying, Luca Ambrosino notes self-conditioning is already standard terminology in diffusion and is becoming the norm for language diffusion.
More from Research
- XPENG's AUGUstus VLA model hits SOTA on NAVSIM and deploys on real vehicles at 12Hz — XPENG-AI · 2026-10-14
- Alibaba study finds most fancy contrastive tricks barely help multimodal embeddings; combos gain 6% — alibaba-multimodal-industrial-ai · 2026-10-14
- Lightricks' APEX pairs hierarchical latents with async denoising for extreme 64-128x diffusion compression — Lightricks · 2026-10-14
- EMHO: Self-Evolving Agent Harness Boosts Frozen Embodied Models via Experience Traces — Hyun Jung Lee · 2026-10-14
- WorldCast: A Distributed Multiplayer World Model That Runs Counter-Strike 2 in Real Time — CUHKSZ · 2026-10-14
- MixLID: A Token-Level Language ID Model Covering 111 Language-Script Combinations — Seongtae Hong · 2026-10-14