PaperlayerView exact revision on arXiv

Source-linked AI summary

Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation

Inferix Team, Tianyu Feng, Yizeng Han, Jiahao He, Yuanyu He, Xi Lin, Teng Liu, Hanfeng Lu, Jiasheng Tang, Wei Wang, Zhiyuan Wang, Jichao Wu, Mingyang Yang, Yinghao Yu, Zeyu Zhang, Bohan Zhuang

arXiv:2511.20714v2cs.CVcs.AI

TL;DR

Long-form world simulation puts pressure on storage and computation, including the GPU memory needed for KV-cache context. Inferix is a block-diffusion inference engine that integrates model pipelines and long-video evaluation; the paper reports unified interfaces and efficiency techniques, while further inference techniques remain future work.

  • Problem

    Long video sequences and large world models pressure storage and computation, while retaining prior-block KV caches consumes substantial GPU memory.

    Source: §2 · p. 2

  • Method

    Inferix provides a block-diffusion inference engine with a generate-and-cache loop, unified interfaces for diverse models, and integrated long-video evaluation.

    Source: §1 · p. 1, §3.3 · p. 4, §6 · p. 6

  • Results

    Inferix unifies inference interfaces for different block-diffusion models and applies efficient inference techniques, while integrating InterVBench to evaluate long-video generation.

    Source: §6 · p. 6

  • Takeaways & Limitations

    The authors present Inferix and InterVBench as tools intended to make world-model and long-video-generation research more convenient and evaluation more precise and efficient.

    Source: §6 · p. 6

    More block-diffusion-specific techniques, including sparse attention, feature cache, and step distillation, remain future work.

    Source: §6 · p. 6

Abstract

from arXiv · show

World models serve as core simulators for fields such as agentic AI, embodied AI, and gaming, capable of generating long, physically realistic, and interactive high-quality videos. Moreover, scaling these models could unlock emergent capabilities in visual perception, understanding, and reasoning, paving the way for a new paradigm that moves beyond current LLM-centric vision foundation models. A key breakthrough empowering them is the semi-autoregressive (block-diffusion) decoding paradigm, which merges the strengths of diffusion and autoregressive methods by generating video tokens in block-applying diffusion within each block while conditioning on previous ones, resulting in more coherent and stable video sequences. Crucially, it overcomes limitations of standard video diffusion by reintroducing LLM-style KV Cache management, enabling efficient, variable-length, and high-quality generation. Therefore, Inferix is specifically designed as a next-generation inference engine to enable immersive world synthesis through optimized semi-autoregressive decoding processes. This dedicated focus on world simulation distinctly sets it apart from systems engineered for high-concurrency scenarios (like vLLM or SGLang) and from classic video diffusion models (such as xDiTs). Inferix further enhances its offering with interactive video streaming and profiling, enabling real-time interaction and realistic simulation to accurately model world dynamics. Additionally, it supports efficient benchmarking through seamless integration of LV-Bench, a new fine-grained evaluation benchmark tailored for minute-long video generation scenarios. We hope the community will work together to advance Inferix and foster world model exploration.

1 Introduction

Block diffusion combines diffusion-style generation within blocks with autoregressive conditioning across blocks, supporting efficient, variable-length video generation. Inferix is built around this paradigm for world simulation, with streaming and long-video evaluation features.

1 Introduction

  • Block diffusion combines diffusion’s parallel generation with AR-style KV caching, enabling efficient, variable-length generation where the alternatives have fixed-length or non-parallel decoding limits.Current video diffusion uses bidirectional attention without KV caching, while AR-based frameworks support variable lengths and caching but have lower generation quality than video diffusion.Source: §1 · p. 1, Figure 1 · p. 2
  • Inferix generates each video block from noise through iterative denoising, using prior blocks’ cached context and updating the cache after each block.This generate-and-cache loop facilitates efficient, arbitrary-length video generation.Source: §1 · p. 1
  • Inferix is purpose-built as a next-generation inference engine for immersive world synthesis using optimized semi-autoregressive decoding.The paper positions it as dedicated infrastructure for the world-model era, distinct from engines associated with LLMs, visual diffusion, and post-training.Source: §1 · pp. 1–2
  • Inferix integrates InterVBench, a fine-grained benchmark for minute-long videos with metrics for long-range coherence.It also provides basic video streaming over RTMP and WebRTC.Source: §1 · p. 2
  • Inferix lists persistent-simulation KV-cache management, distributed world synthesis, built-in diffusion profiling, and segment-specific prompt control among its features.Different prompts can control different video segments.Source: §1 · p. 2

2 Challenges in Inference of World Simulation

World simulation combines long video sequences with large models, creating storage and computation pressure. The section identifies KV-cache management and inference acceleration as key challenges.

2 Challenges in Inference of World Simulation

  • KV caches are a storage bottleneck because retaining prior-block context for future video generation consumes substantial GPU memory.The paper says this context helps ease drift and forgetting, and identifies PageAttention, offloading, and KV-cache compression as techniques to bring from LLM inference.Source: §2 · p. 2
  • Long sequences and large models greatly increase computation; generating a 5-second video with Wan2.1 14B on one NVIDIA H20 takes about 6,800 seconds.Longer context makes world-simulation computation heavier still.Source: §2 · pp. 2–3
  • The framework groups efficiency components around parallel strategies, block-wise KV management, DAX quantization, streaming, and fine-grained evaluation.The paper lists low-bit quantization, sparse attention, fewer denoising steps, inference redundancy, and distributed computation as possible acceleration approaches.Source: Figure 2 · p. 3, §2 · p. 3

3 Framework Design

Inferix combines parallelism, block-wise KV management, model-specific pipeline abstractions, profiling, and video-streaming support. The design targets varied block-diffusion models and long-sequence inference.

3.1 Parallelism

  • Inferix uses Ulysses-style sequence parallelism and Ring Attention, choosing a strategy based on model architecture, network topology, and communication overhead.Ulysses partitions attention heads across GPUs; Ring Attention distributes attention computation in a ring topology.Source: §3.1 · p. 3

3.2 KV Management

  • A unified KV-management interface supports block-wise memory, flexible range- or index-based fetches, MLA latent storage, and offloading to main memory.The design accommodates sliding-window access and selective global context dependencies.Source: §3.2 · p. 3

3.3 Models and Pipelines

  • Inferix abstracts shared computation across MAGI-1, CausVid, and Self Forcing into a generalized pipeline with extensible interfaces for user-integrated models.CausVid and Self Forcing build on Wan2.1, while MAGI-1 is trained from scratch with distinct infrastructure.Source: §3.3 · p. 4

3.4 System Profiling

  • The built-in profiler provides end-to-end resource visibility with less than 5% overhead and supports custom metrics through hooks or callbacks.Users can instrument functions with a Python decorator or broader code regions with a context manager.Source: §3.4 · p. 4

3.5 Video Streaming

  • For CausVid, users can change prompts between video chunks, and Inferix clears the cross-attention cache when a new prompt is supplied.The section motivates dynamic narrative control with different signals for different chunks.Source: §3.5 · p. 4

4 InterVBench

InterVBench targets long-form video evaluation with a curated dataset and metrics for temporal drift and video quality. Its data-engine prompt asks for coherent, detailed descriptions of successive frames.

4.1 Dataset

  • InterVBench comprises 1,000 high-resolution videos exceeding 50 seconds, sourced from DanceTrack, GOT-10k, HD-VILA-100M, and ShareGPT4V.Captions are generated every 2–3 seconds, checked through human review at sourcing, segmentation, and caption verification stages, and the curated data is split 80/20 into train and evaluation sets.Source: §4.1 · pp. 4–5

4.2 Metrics

  • Video Drift Error measures relative quality changes over time, and its five variants assess clarity, motion, aesthetic consistency, background stability, and subject identity drift.Lower VDE scores indicate stronger temporal consistency.Source: §4.2 · p. 5
  • InterVBench also incorporates five VBench quality dimensions: subject and background consistency, motion smoothness, aesthetic quality, and image quality.Higher scores indicate better performance on these dimensions.Source: §4.2 · p. 5

4.3 Prompts for InterVBench’s Data Engine

  • The data-engine prompt asks for one coherent English paragraph describing a frame’s subject, actions, expressions, setting, lighting, and cinematic qualities.It supplies the previous frame’s description as context to support temporal coherence.Source: §4.3 · p. 5

5 Development Roadmap

The roadmap lists planned capabilities for Inferix spanning KV management, model adaptation, deployment, and streaming.

  • Planned capabilities include flexible block-sparse KV management, finetuning pretrained video models into semi-AR, and distilling models into few steps.Source: §5 · p. 5
  • The roadmap also includes support for high-concurrency deployment and more complex distributed inference.Source: §5 · p. 5
  • Streaming plans include improving usage and performance, alongside more advanced real-time interactive capabilities.Source: §5 · p. 5

6 Conclusion

Inferix is presented as a block-diffusion inference engine with unified interfaces, efficiency techniques, and an integrated long-video benchmark. The conclusion identifies more block-diffusion-specific efficiency techniques as future work.

  • Inferix unifies interfaces for block-diffusion models and applies efficient inference techniques to improve ease of use.Source: §6 · p. 6
  • Integrated InterVBench aims to evaluate long-video generation quality precisely and efficiently, supporting world-model development.Source: §6 · p. 6
  • Future work will consider block-diffusion-specific efficiency techniques, including sparse attention, feature caching, and step distillation.Source: §6 · p. 6

A Contributions and Acknowledgments

The contributors describe a joint team affiliated with Zhejiang University, Hong Kong University of Science and Technology, and Alibaba, and invite others to join the virtual team.

Loading 2511.20714v2…