Video Editing Frontiers — Topic Summary

Successful editing requires the requested change, preservation of unrelated content, and a coherent event over time.

Method landscape

The presentation compares inference-time transport, learned instruction editors, local and compositional controls, motion and action editing, counterfactual changes, long and multi-shot video, multi-turn history, and tool-driven systems. Dedicated pages cover FlowEdit (ICCV 2025), DynaEdit, Bernini, MotionV2V, and other representative methods.

Mechanisms and evidence

FlowEdit supplies an image-editing transport formulation; video adaptation requires a video backbone. DynaEdit targets dynamic rewriting through its source queries and guidance. Bernini separates semantic planning from video rendering. These routes use different conditioning signals and training assumptions, so their capabilities need task-specific evidence.

World-model approaches introduce visual state, geometry, and interventions. World in World and NeoVerse provide different uses of source evidence. Moving from camera or appearance changes to existing-object interaction requires additional state recovery and contact constraints. ActionSplice intervenes during generation; transferring that mechanism to arbitrary recorded video needs source alignment.

Aurora and MMLVE-Agent motivate tool preparation, shot coordination, memory, and feedback. Their end-to-end gains should be compared with fixed workflows and repeated sampling under matched budgets.

Research questions

Central claim: Editing success must be checked at the level of objects, events, and valid edit history, alongside visual quality.

Hidden assumption: A desired action visible in the output may involve a new object, omit the required transition, or retain the event that should have been replaced.

Failure boundary: Existing-object contact, action substitution, occlusion, history-dependent instructions, and judge protocol changes expose distinct weaknesses.

Connection: Causal video editing links a requested intervention to its necessary consequences while retaining unrelated source evidence.

Concrete experiment: On the same source cases, compare outputs with no edit, an object substitute, an appended action, and a completed interaction; calibrate the judge against explicit object and event annotations.

Discussion seed: Which intermediate representation makes object interaction both controllable and verifiable?

Open questions

How can an editor operate on the same object already present in the source, with continuous contact and state change?
Which history should be stored to support instruction updates, selective undo, and reappearance after occlusion?
How can world models recover uncertain source state, propagate intervention consequences, and correct discrepancies in rendered video?
When do planning, tool choice, and local repair improve an editing agent under a fixed compute budget?
How should benchmarks expose wrong objects, missing contact, retained actions, and forgotten history without rewarding no-edit outputs?

These are research questions and candidate experiments. The related methods supply specific mechanisms; the proposed combined system has not been established by the examples.

Examples and protocol notes

The deck includes action substitution, clothing edits across turns, and a dog–bone interaction comparison. They illustrate failure types without estimating a model-wide rate. The evaluation discussion uses a frozen 2026-09-26 snapshot. Read the evidence notes.