MEVG: Multi-event Video Generation with Text-to-Video Models

TL;DR

We introduce MEVG, a training-free diffusion-based method that generates a single coherent video depicting multiple sequential events given individual text sentences — with no fine-tuning or additional video data required. A Last Frame-aware Latent Initialization (LFLI) strategy seeds each clip's latent from the previous clip's final frame, while Dynamic Noise enforces diversity and Structure-guided Sampling maintains visual consistency within each event. A novel prompt generator automatically converts coarse user scenarios into optimized per-event prompts, and MEVG outperforms zero-shot baselines in multi-event video generation.

Key Contributions

• Training-free multi-event generation — generates a video consisting of multiple events without requiring any training or additional video data.
• Last-frame-aware initialization & dynamic noise adjustment — a novel latent vector strategy enhancing temporal and semantic consistency between consecutive event clips.
• Novel prompt generator — transforms coarse text inputs into optimal per-event instructions, ensuring coherent semantic transitions throughout the generated video.
• State-of-the-art zero-shot performance — outperforms other zero-shot video generation methods in multi-event reflection while maintaining visually coherent content.

Project Design

Given multiple prompts (P) each describing a different event, MEVG chains video clips into a single, temporally coherent video. Each clip's latent is initialized from the final frame of the preceding clip via Last Frame-aware Latent Initialization (LFLI), and Structure-guided Sampling ensures visual consistency throughout each event.

(i) Structure-guided Sampling

• Maintain visual consistency within an event, which act as a regularization term.
• Differential update strategy prevent the identical latent codes across the video.

(ii) Dynamic Noise

• Enforce the diversity of the generated video with noise scheduling function.
• Noise scheduling function is monotonically decrease function.

(iii) Last Frame Aware Inversion

• Maintain a visual correlation between two different events.
• Denoised observation x̂ contains a sketcy spatial layout and context.

Results

Applications

"A single white flower gradually blooms from a single green flower bud."

→"The single white flower is blooming."

→"A lovely fully blossomed single white flower."

Input image

"People walks on the beach at night."

→"There are sand castles on the beach under the fireworks at night."

→"Very few people remain on the beach at night and they gradually fade away."

Input image

←

→

Conclusion

We presented MEVG, a training-free diffusion-based framework for generating temporally coherent videos from multiple sequential text prompts — requiring no fine-tuning or additional video data. By combining Last Frame-aware Latent Initialization, Dynamic Noise scheduling, and Structure-guided Sampling, MEVG effectively bridges consecutive event clips while maintaining visual and semantic consistency. A novel prompt generator further enables flexible coarse-to-optimal text input. Extensive experiments and user studies demonstrate that MEVG outperforms zero-shot baselines in temporal coherency of content and semantics, offering a practical and scalable approach to multi-event video synthesis.