MoviEdit: Long-form Multi-Shot Video Editing Made Easy

MoviEdit is an end-to-end framework for long-form multi-shot video editing that combines scalable paired-data generation, a cascaded bidirectional–autoregressive editor, and a dedicated benchmark.

Long-form multi-shot video editing requires an intended edit to remain coherent across cuts in scene, viewpoint, and subject appearance while processing much longer visual context than single-shot editors. Existing video editing methods remain predominantly designed for short, single-shot videos. The paper identifies coupled bottlenecks in data, modeling, and evaluation: cross-shot paired supervision does not exist at scale, joint processing of full source and target videos becomes computationally prohibitive, and existing benchmarks do not assess cross-shot localization and consistency.

MoviEdit builds three annotation-free data-generation pipelines: pixel-aligned editing, subject localization, and semi-pixel-aligned editing, producing paired multi-shot editing videos with cross-shot subject correspondence. Pixel-aligned editing decomposes multi-shot editing into joint first-frame image-grid editing across shots and within-shot temporal propagation using a first-frame-conditioned single-shot video editor. Semi-pixel-aligned editing couples source and target generation by sharing per-shot captions and initial noise, with task-specific inference-time interaction and optional ControlNet conditioning. The model uses a cascaded architecture with shared weights: a lightweight bidirectional Draft Editor performs full-sequence diffusion on a spatiotemporally subsampled video, and a causal Render Editor autoregressively propagates and refines the edit chunk by chunk at full resolution conditioned on the draft. Inference is accelerated with velocity moment matching distillation and KV caching, and evaluation is performed with MoviEdit-Bench.

• The data pipeline constructs 87K paired multi-shot editing videos from 61,251 raw clips; the detailed table reports 205,228 raw pairs and 86,619 filtered pairs at 540p, 15s, and 24 fps. §4.1
• Quality filtering rejects 17%/38% of multi-camera and 32%/47% of cinematic pixel-aligned pairs at the two stages, 24% of localization pairs, and 66% of semi-pixel-aligned pairs. §3.1
• For text-guided V2V editing on MoviEdit-Bench, MoviEdit achieves overall edit success of 0.790, compared with 0.764 for Aleph 2.0 and 0.378 for Wan-Edit. §4.2
• For edited-frame-guided IV2V editing, MoviEdit reaches 0.803 overall edit success on the first-frame subset and 0.786 on arbitrary frames, compared with Aleph 2.0 at 0.806 and 0.781, respectively. §4.2
• At 540p and 361 frames on 8 H100 GPUs, denoising latency falls from 364.1 s with full temporal attention to 258.9 s with sliding-window attention, 166.8 s with KV caching, and 11.8 s with distillation, a 30.9× speedup. §4.2
• Scaling either training duration or dataset size consistently improves performance; the study varies 0%, 20%, 40%, 80%, and 100% of the 17K-step schedule and 0%, 33%, and 100% of the 87K-pair dataset. §4.2

• The paper states that target localization can fail under strong identity ambiguity, such as when multiple visually similar subjects are present, causing the edit to be applied to the wrong target.
• The paper states that partial target visibility, such as only a body part or object fragment being visible, can make cross-shot association difficult and lead to incomplete or incorrect edits.
• MoviEdit focuses on correspondence-preserving edits that retain source motion and narrative structure; extending to free-form transformations of motion and narrative structure is stated as future work.
• Beyond-training-horizon results are qualitative on 25-second videos, and MoviEdit-Bench uses 64 generated source videos and 128 editing scenarios, leaving broader quantitative validation open.

• Three annotation-free pipelines provide supervision: pixel-aligned editing, subject localization, and semi-pixel-aligned editing, leveraging synchronized multi-camera footage, cinematic clips, and generated multi-shot content. §3.1
• Pixel-aligned editing edits the first frames of all shots jointly as an image grid, then propagates each edited first frame within its shot using a pretrained first-frame-conditioned single-shot video editor. §3.1
• Semi-pixel-aligned editing couples source and target generations by sharing per-shot captions and initial noise, uses a fork timestep for character appearance change, and uses HED or depth ControlNet conditioning for environment and medium change. §3.1
• The cascaded architecture uses a bidirectional Draft Editor over a subsampled sequence, empirically M=N/2, to establish global edit coherence, followed by a causal Render Editor that processes full-resolution chunks conditioned on the draft. §3.2
• The Draft Editor and Render Editor share model weights and differ in attention patterns and spatiotemporal resolution; training uses flow matching for the Draft Editor and a causal diffusion forcing objective for the Render Editor. §3.2
• Inference optimization uses velocity moment matching distillation with asymmetric schedules, distilling the Draft Editor to 12 steps and the Render Editor to 4 steps, plus KV caching for reusable tokens. §3.3
• MoviEdit-Bench contains 64 generated multi-shot videos and 128 editing scenarios across eight categories, supporting V2V and IV2V with first-frame and non-first-frame references and local and global edits. §3.4
• The paper reports generalization to multi-reference-frame editing, video set-to-set editing, and beyond-training-horizon inputs of 25 seconds with up to 17 shots, although trained on videos up to 15 seconds. §4.2