This is a repo to track the latest autoregressive visual generation papers.

Sekai2: From World Exploration to Interactive World Modeling (arXiv:2608.09449) — 128,892 real-world clips with camera trajectories and temporally grounded captions.

On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

Open Frontier Intelligence

Vidu S1: A Real-Time Interactive Video Generation Model

ICLR'26: TTOM: Test-Time Optimization and Memorization for Compositional Video Generation

[ICML 2026] The official implementation of paper "Unified Multimodal Autoregressive Modeling with Shared Context—Visual Tokenizer is Key to Unification"

Official codebase for "Self Forcing: Bridging Training and Inference in Autoregressive Video Diffusion" (NeurIPS 2025 Spotlight)

code for "Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion"

Code for RepWAM: World Action Modeling with Representation Visual-Action Tokenizers

Next Forcing: Causal World Modeling with Multi-Chunk Prediction (MCP)

ARM: An AutoRegressive Large Multimodal Model with Discrete Representations

[CVPR 2026 Best Paper Finalist] Pixel Diffusion Transformers for Image Generation

[SIGGRAPH Asia 2026] DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models

PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion

Flow Map OPD for AnyStep Video Diffusion

[ICML 2026] RoboTwin 2.0 Offical Code Repo

Benchmarking Knowledge Transfer in Lifelong Robot Learning

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models

[ICLR 2026 Oral] DiffusionNFT: Online Diffusion Reinforcement with Forward Process

[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.

Official Repo of "D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models"

slime is an LLM post-training framework for RL Scaling.

[Nips 2025] EgoVid-5M: A Large-Scale Video-Action Dataset for Egocentric Video Generation

Code accompanying Ego-Exo: Transferring Visual Representations from Third-person to First-person Videos (CVPR 2021)

HY-SOAR:Self-Correction for Optimal Alignment and Refinement in Diffusion Models

Official implementation of "OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning"

Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models