Unlocking any-order generation in pretrained autoregressive image models with future-aware embeddings
Rishav Pramanik · hf · 2026-10-31
A Stony Brook University / ÉTS Montréal (LIVIA) paper shows that pretrained raster-scan autoregressive image models can be adapted into flexible any-order generators via future-aware positional embeddings and an ordering curriculum, enabling inpainting and outpainting with minimal extra compute. The authors argue this beats discrete-diffusion alternatives that train from scratch on mixed token orderings at roughly 3× the cost and sacrifice compatibility with the AR inference ecosystem (KV caching, vLLM): the pretrained AR model has already learned the joint distribution over image tokens — it only needs the ability to express it under non-sequential conditioning.
More from Multimodal
- Image-understanding model Jev 1.14 lands on Runware at $0.042 per 1M input tokens — aziz4ai · 2026-10-31
- ComfyUI workflow turns MiniMax H3 into a full comic-page generator, open-sourced — gatortux · 2026-10-31
- Runway Agent gains new skills: opens files, runs code, builds brand launch kits — runwayml · 2026-10-31
- FastVideo open-sources FastH3: 4-step distilled MiniMax H3 with 14x Blackwell speedup — victormustar · 2026-10-31
- AI-generated series 'Árpád' hits episode 6 with strikingly lifelike results — PannoniumFilms · 2026-10-31
- Deconstruct AI reverse-engineers reference images into MJ/SDXL/ComfyUI prompts — robert-988 · 2026-10-31