AI Paper Atlas

Research map / Video / Video generation

HorizonComposer: Spatiotemporally Consistent Driving Video Editing with Enriched Traffic Semantics

Mauricio Soroco, Yuqiu Liu, Zaid Tasneem, Francesco Pittaluga, Abhishek Aich, Wuyang Chen, Manmohan Chandraker, Ziyu Jiang

NeurIPS 2026 · Video › Video generation · approach: World-model video generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HorizonComposer, an instruction-guided driving video editing method that preserves traffic semantics and spatiotemporal consistency when simulating diverse driving conditions.

Problem

General-purpose instruction-guided video editors corrupt traffic semantics when applied to global driving-condition edits, which undermines their use for stress-testing autonomous driving systems within the Operational Design Domain (ODD).

Approach

Tailors instruction-guided video editing to driving scenes by enriching traffic semantics so that global condition edits (e.g., weather, lighting) keep traffic elements consistent across frames. Only the abstract was available, so finer method details are not stated.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add dense pixel-level reward conditioning (ablation)HorizonComposer · OOD (nuPlan)Lane IoU: 0.5984 -> 0.6266Improved
add dense pixel-level reward conditioning (ablation)HorizonComposer · OOD (nuPlan)Car appearance: 0.8308 -> 0.8595Improved
add dense pixel-level reward conditioning (ablation)HorizonComposer · In-domainLane IoU: 0.4276 -> 0.4904Improved
add thinking frames (temporal augmentation) (ablation)HorizonComposer · OOD (nuPlan)FF-alpha: 3.3927 -> 3.2185Improved
add thinking frames (ablation)HorizonComposer · OOD (nuPlan)Subject consistency: 0.9011 -> 0.9126Improved
add thinking frames (ablation)HorizonComposer · OOD (nuPlan)instruction success rate: 0.9048 -> 0.8762Hurt
add depth conditioning to RGB-only fine-tune (ablation)HorizonComposer · In-domainFF-alpha: 3.6594 -> 3.4901Improved
add dense rewards on top of depth conditioning (ablation)HorizonComposer · In-domaininstruction success rate: 0.9489 -> 0.9333Hurt
plain RGB-only fine-tuning versus synthetic pipeline (ablation)HorizonComposer · OOD (nuPlan)Lane IoU: 0.5774 -> 0.6006No clear effect
HorizonComposer versus SyntheticPip on lane segmentation downstreamPIDNet training data · SeRMmIoU (%): 38.45 -> 40.27Improved
add generated videos to real BDD100K training datadetector fine-tuning · BDD100KmAP: 0.167 -> 0.175Improved
HorizonComposer versus SyntheticPip on instruction alignment and image qualityHorizonComposer · In-domainrelative change vs SyntheticPip: Succ. -1.34, Img. -4.57Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Baselines: general-purpose instruction-guided video editors

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy