Alec Radford tests a GPT‑style model trained only on physics literature up to 1911, asking it to independently infer the mathematical form of Einstein’s general relativity, thereby probing the computability of scientific discovery and the creative reasoning limits of current AI systems.
VA‑π introduces a lightweight post‑training framework that uses variational inference and reinforcement learning to align tokenizers with visual autoregressive generators, achieving dramatic quality gains, extreme training efficiency, and robust pixel‑level reconstruction across diverse image generation tasks.
A comprehensive evaluation shows that while diffusion language models achieve higher generation speed through parallel decoding, they suffer from severe causal reasoning and formatting deficiencies, lagging far behind autoregressive models on embodied and tool‑calling agent tasks.
The article examines the current dominance of diffusion models in commercial video generation, contrasts them with autoregressive methods, and details how the open‑source MAGI‑1 model combines both paradigms to achieve longer, more controllable video synthesis while addressing scalability and quality challenges.
The article analyzes the Token‑Shuffle method, which reduces visual token redundancy to allow high‑resolution (2048×2048) autoregressive image generation, detailing its architecture, training pipeline, experimental results, efficiency gains, and comparisons with diffusion and other AR models.
This comprehensive survey reviews the rapid progress of multimodal understanding and text‑to‑image generation models, categorises existing unified architectures into diffusion‑based, autoregressive, and hybrid paradigms, analyses their tokenisation strategies, datasets and benchmarks, and highlights current challenges and future research directions.
The paper presents FAR, a frequency‑aware autoregressive framework that predicts image tokens from low‑frequency to high‑frequency components using a continuous tokenizer, and demonstrates its efficiency and quality on ImageNet and text‑to‑image benchmarks compared with existing AR and VAR methods.
FlexVAR replaces residual prediction with direct ground‑truth prediction in visual autoregressive modeling, enabling generation of arbitrary resolutions and aspect ratios, supporting image‑to‑image tasks such as inpainting and upscaling, and offering adjustable inference steps that trade speed for quality while achieving state‑of‑the‑art FID scores.
The paper proposes a Next Patch Prediction (NPP) paradigm that groups image tokens into high‑density patches, enabling autoregressive models to predict patches instead of individual tokens, which reduces training cost to about 0.6× and improves ImageNet FID scores by up to 1.0 across models ranging from 100 M to 1.4 B parameters.
The VAR model, a Visual AutoRegressive framework that introduces a novel multi‑scale “next‑scale prediction” paradigm, dramatically improves image generation efficiency and quality, surpasses diffusion models, validates scaling laws in vision, and earned the Best Paper award at NeurIPS 2024.
