SwitchCraft: Training-Free Multi-Event Video Generation with Attention Controls

Qianxun Xu; Chenxi Song; Yujun Cai; Chi Zhang

cs.CVFeb 27, 2026

SwitchCraft: Training-Free Multi-Event Video Generation with Attention Controls

Authors: Qianxun Xu, Chenxi Song, Yujun Cai, Chi Zhang

Organizations: Westlake University · Duke Kunshan University · The University of Queensland

Abstract

Recent advances in text-to-video diffusion models have enabled high-fidelity and temporally coherent videos synthesis. However, current models are predominantly optimized for single-event generation. When handling multi-event prompts, without explicit temporal grounding, such models often produce blended or collapsed scenes that break the intended narrative. To address this limitation, we present SwitchCraft, a training-free framework for multi-event video generation. Our key insight is that uniform prompt injection across time ignores the correspondence between events and frames. To this end, we introduce Event-Aligned Query Steering (EAQS), which steers frame-level attention to align with relevant event prompts. Furthermore, we propose Auto-Balance Strength Solver (ABSS), which adaptively balances steering strength to preserve temporal consistency and visual fidelity. Extensive experiments demonstrate that SwitchCraft substantially improves prompt alignment, event clarity, and scene consistency compared with existing baselines, offering a simple yet effective solution for multi-event video generation.

Figures & tables

Figure 2: Overview of SwitchCraft. (a) EAQS takes a text prompt and user specified event time spans, identifies anchor tokens for each event, and constructs event specific projectors from their attention keys. It steers video queries toward the target event and away from others in each temporal span. (b) ABSS estimates enhancement and suppression strengths by extracting dominant directions from the event keys and correcting the attention deficit. The updated queries pass through the video diffusion transformer so that each temporal span follows its intended event with smooth transitions.

Figure 3: Qualitative comparison. SwitchCraft executes all events in the intended order, prevents leakage and omission, and maintains subject and scene consistency. Baselines show omissions, drift across spans, or progressive quality decay.

	CLIP [ 31 , 13 ]		VideoScore2 [ 11 ]			VBench [ 17 ]
Method	CLIP-T $\uparrow$	CLIP-F $\uparrow$	Visual quality $\uparrow$	T2V alignment $\uparrow$	Phys. cons. $\uparrow$	Motion smoothness $\uparrow$	Sub cons. $\uparrow$	Back cons. $\uparrow$	Aesthetic quality $\uparrow$	Imaging quality $\uparrow$
MEVG [ 28 ]	0.244	0.915	2.13	2.33	1.73	0.953	0.701	0.841	0.346	0.525
DiTCtrl [ 3 ]	0.246	0.959	3.20	3.27	2.93	0.981	0.764	0.876	0.511	0.702
LongLive [ 40 ]	0.252	0.984	4.27	3.13	3.97	0.984	0.898	0.908	0.627	0.725
Wan2.1 [ 37 ]	0.256	0.980	4.30	3.47	4.12	0.987	0.947	0.924	0.645	0.738
Stitch	0.257	0.963	3.73	3.67	3.80	0.983	0.926	0.910	0.608	0.711

Table 1: Quantitative comparison and ablation study. SwitchCraft improves text alignment and preserves temporal smoothness and visual quality across CLIP [ 31 , 13 ] , VideoScore2 [ 11 ] , and VBench [ 17 ] . Best scores are in bold and second best are underlined . Columns highlighted in yellow correspond to text alignment metrics, purple to temporal smoothness metrics, and blue to visual quality and consistency metrics. Removing components in the ablations generally reduces text alignment, while the other metrics remain close to the Wan 2.1 [ 37 ] backbone.

Figure 4: Failure of Stitch . Segment-wise generation inherits bias from the last frame of the previous segment and propagates motion and layout, causing action bleed and unstable handoffs.

Figure 5: More applications. SwitchCraft generates creative occluding transitions between scenes, preventing event bleeding while preserving identity and global context.

Method	No omission $\uparrow$	No leakage $\uparrow$	Smooth transition $\uparrow$	Visual quality $\uparrow$
MEVG [ 28 ]	1.41	1.38	1.38	1.28
DiTCtrl [ 3 ]	1.66	1.48	1.48	1.59
LongLive [ 40 ]	2.07	2.72	2.97	3.52
MinT [ 39 ]	4.31	3.69	3.76	3.83
Wan 2.1 [ 37 ]	3.17	3.38	3.79	3.93
Stitch	2.62	2.07	2.14	2.45

Table 2: Human evaluation. Users scored no omission, no leakage, transition smoothness, and visual quality on a five point scale. SwitchCraft attains the best overall performance.

Figure 6: Ablation study of key components. Prompt: On a snowy plain, a dog is running forward, then suddenly stops to sniff the ground, then continues running. The full method yields complete sequences with stable quality. Removing parts leads to event leakage, omission, and degraded visual quality.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
DiT blocks	$40$
Attention heads	$40$
Video frames	$81$
Sampler	UniPC
Guidance scale	$5.0$

Appendix

Table 3: Hyperparameters for the Wan 2.1 T2V 14B [ 37 ] backbone used by SwitchCraft.

Figure 7: Limitations. (1) Complex actions: When the underlying architecture struggles with specific movements, the output defaults to basic approximations. (2) Multi-subject: Our method does not have spatial constraints for solving identity-action entanglement between characters.

Figure 8: Comparison of creative occluding transitions. SwitchCraft produces in-shot occluding transitions that blend the first and last scenes while preserving subject identity, whereas baselines leave residual transition elements or degrade identity and scene coherence.

Figure 9: Comparison of creative occluding transitions. SwitchCraft produces in-shot occluding transitions that blend the first and last scenes while preserving subject identity, whereas baselines leave residual transition elements or degrade identity and scene coherence.

Figure 10: Additional Qualitative comparison. SwitchCraft executes all events in the intended order, prevents leakage and omission, and maintains subject and scene consistency. Baselines show omissions, drift across spans, or progressive quality decay.

Figure 11: Additional qualitative results. For diverse multi-event prompts, SwitchCraft generates temporally coherent videos with smooth transitions between events, while preserving subject identity and global context.

Number of events	Wan 2.1 [ 37 ]	SwitchCraft
2	15.2	17.6
3	15.2	19.7
4	15.2	22.3

Appendix

Table 4: Inference time per video (minutes) for Wan 2.1 T2V 14B [ 37 ] and SwitchCraft on multi-event prompts of different lengths.

	2 events	3 events	4 events
Base step (no edit)		18.66
EAQS step (fixed $\alpha,\beta$ )	19.62	20.07	20.81
EAQS+ABSS step	22.32	26.83	30.19

Appendix

Table 5: Inference time per step (seconds) for the base model, EAQS, and EAQS+ABSS on multi-event prompts of different lengths.

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

Prompt Relay: Inference-Time Temporal Control for Multi-Event Video Generation
Apr 11, 2026Gordon Chen, Ziqi Huang, Ziwei LiuVideo Generation Controllable Video Generation
S-Lab, Nanyang Technological University
TS-Attn: Temporal-wise Separable Attention for Multi-Event Video Generation
Apr 21, 2026Hongyu Zhang, Yufan Deng, Zilin Pan +6Image-to-Video Generation T2V Generation
Peking University, Shenzhen Graduate School · Zhejiang University · Nanjing University +3
TunerDiT: Training-free Progressive Steering of Diffusion Transformer for Multi-Event Video Generation
May 29, 2026Ruotong Liao, Guowen Huang, Qing Cheng +6Diffusion Transformer T2V Generation
Ludwig Maximilian University of Munich · MCML · Technical University of Munich +2

Apr 11, 2026J/K move · Enter open · S save

Prompt Relay: Inference-Time Temporal Control for Multi-Event Video Generation

Gordon Chen, Ziqi Huang, Ziwei Liu

S-Lab, Nanyang Technological University

Video Generation Controllable Video Generation Video Diffusion Models Attention Control in Diffusion Models

Open paper AskPDF arXiv

Video diffusion models have achieved remarkable progress in generating high-quality videos. However, these models struggle to represent the temporal succession of multiple events in real-world videos and lack explicit mechanisms to control when semantic concepts appear, how long they persist, and the order in which multiple events occur. Such control is especially important for movie-grade video synthesis, where coherent storytelling depends on precise timing, duration, and transitions between events. When using a single paragraph-style prompt to describe a sequence of complex events, models often exhibit semantic entanglement, where concepts intended for different moments in the video bleed into one another, resulting in poor text-video alignment. To address these limitations, we propose Prompt Relay, an inference-time, plug-and-play method to enable fine-grained temporal control in multi-event video generation, requiring no architectural modifications and no additional computational overhead. Prompt Relay introduces a penalty into the cross-attention mechanism, so that each temporal segment attends only to its assigned prompt, allowing the model to represent one semantic concept at a time and thereby improving temporal prompt alignment, reducing semantic interference, and enhancing visual quality.

Abstract

Video Generation

Explore similar work

Prompt Relay: Inference-Time Temporal Control for Multi-Event Video Generation

TS-Attn: Temporal-wise Separable Attention for Multi-Event Video Generation

TunerDiT: Training-free Progressive Steering of Diffusion Transformer for Multi-Event Video Generation