AIAny
Icon for item

InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter

Continues a live or ongoing video stream while applying user-specified edits on the fly using a lightweight edit-ignition adapter. The adapter injects edits only in chunks where requests arrive and uses history cross-attention and temporal causal self-attention to preserve continuity and stability for unbounded streaming edits.

Introduction

Most video editing methods assume a fixed input clip and perform in-place, frame-aligned edits. That design breaks down for live or open-ended streams where edits must be applied as the stream continues: edits should behave like a faithful continuation, not a frame-by-frame rewrite, and quality must not degrade as edits accumulate. InfinityEdit tackles this operational gap by attaching a minimal adapter to a streaming generator so edits can be ignited at any time while preserving the model's original infinite-generation capability.

Key Findings
  • Lightweight, targeted intervention: the edit-ignition adapter is small and activated only in the chunk where an edit request arrives, so the base streaming generator remains unchanged for subsequent segments — this means edits can be applied without sacrificing the model's long-horizon generation behavior.

  • Three specialized attention paths: history cross-attention guides denoising with prior frames, temporal causal self-attention enforces a strictly forward-flowing temporal signal, and edit cross-attention injects the edit instruction. Together they enable edits that respect prior motion and identity while changing appearance or semantics where requested.

  • Data and stability focus: the paper introduces a data-collection pipeline for the infinite-editing setting and evaluates long sequences of successive edits, demonstrating that the method produces faithful continuations and resists drift as edits accumulate — this addresses a practical failure mode of many per-clip editors.

Who it's for and tradeoffs

Great fit if you need real-time or open-ended editing workflows (live streams, gameplay restyling, ongoing camera shoots) where future frames are unavailable and edits must persist into newly generated content. The approach is also useful when you want minimal changes to an existing streaming generator by attaching a small adapter instead of retraining the full model.

Look elsewhere if you require exact frame-wise retransformation of a fixed clip (offline, bidirectional editors tend to yield higher per-frame fidelity) or if you need a turnkey application with UI and deployment tooling — InfinityEdit focuses on model-level editing semantics and stability rather than product integration.

Where it fits

InfinityEdit sits between offline, bidirectional video editors (which can revisit future frames) and fully unmodified streaming generators. Its niche is causal, chunked generation with editable behaviors: it converts offline edit intents into transient, localized interventions that persist forward without reinterpreting future frames.

Information

  • Websitearxiv.org
  • AuthorsYunze Tong, Mushui Liu, Canyu Zhao, Shiyi Zhang, Didi Zhu, Peng Zhang, Wanggui He, Jinlong Liu, Ying Chen, Hao Jiang
  • Published date2026/08/21

More Items

Benchmarks assistant-style, multi-turn interaction for omni-modal LLMs on real-time video by reverse-engineering Internet clips into guided multi-turn interactions. It provides predefined priors and segment-level constraints so models must follow exact routes while being evaluated on answer correctness, timing, visual-prompt handling, and context retention.

Generates group images that bind up to ten reference identities to distinct people and locations by predicting an explicit identity–layout plan and supervising faces with Layout-Grounded ID Loss. Improves identity fidelity while cutting copy-paste duplication; suited for multi-person image synthesis but requires identity-annotated face regions and paired training data.

Turns an uncalibrated monocular actor video into multiview-consistent novel-view videos and lifts them into 4D Gaussian Splatting assets. Introduces Reference Context Packing to keep reference conditioning fixed-size and Target Context Routing to exchange context across target groups, improving large-view reconstruction consistency.