Researchers at Tele-AI have published ShotPlan, a framework that adds explicit shot planning to video generation models. The problem is concrete: current video generators produce coherent single clips, but if you ask them for a sequence of multiple shots with a flowing narrative, they lose the thread.

The mechanism is simple in concept. A single planning token, replicated for each required transition and positioned at fractional temporal coordinates via FRoPE, tells the pre-trained model where to cut. No architectural changes, no attention masks. Clean cuts land within a frame of the requested index, and the same system handles soft transitions and localized camera movements.

Why it matters to you. If you generate video with AI for work, you know the leap: from “a nice 5-second clip” to “three shots that tell a product video story.” That’s where current tools break. You have to generate each clip separately, edit them by hand, and hope characters and style hold together between shots. ShotPlan takes a step toward controlled multi-shot generation, the format you actually need for ads, social content, educational videos.

The trained models, dataset, and 30+ demos are public on Hugging Face and the project site. The base model is Wan2.2-T2V-A14B, a 14-billion-parameter video diffusion model.

If you want to try it, the weights are on Hugging Face (Pensioner/ShotPlan-Wan2.2-T2V-A14B-HighNoise) along with the training dataset. You’ll need a GPU capable of running a video diffusion model of that scale.

In detail

Diffusion-based video generation models produce a fixed number of frames from a text prompt. They work well for a single shot: a person walking, a landscape moving. But cinema, even the basic kind you need for a product video, is made of cuts. A wide shot first, then a detail, then a pan. If you try to describe it all in one prompt, the model mixes the shots or ignores them.

What came before. Anyone wanting to produce multi-shot sequences with AI used two strategies, both unsatisfying. The first: generate separate clips and edit them together, with the problem that characters and environments aren’t consistent between clips. The second: use extremely long prompts hoping the model respects a storyboard, with unpredictable results. Some academic approaches tried conditioning the model with attention masks or modified architectures, but they required training from scratch or heavy changes to the base model.

What changes. ShotPlan introduces a planning token that works as a temporal marker. The token is inserted in the input sequence and linked to a precise temporal coordinate through FRoPE (Fractional Temporal Rotary Position Embedding). Essentially, it tells the model: “at frame 47, change shot.” The model learns to respect these markers during training, without changing its base architecture.

The result, according to the authors, is that clean cuts occur within a frame of the requested index. The same mechanism supports dissolves and camera movements limited to a temporal portion of the clip. This means you can specify a shot sequence in the prompt and the model respects it at the right frames.

The numbers. The authors claim ShotPlan outperforms existing methods for inter-shot coherence and shot handling flexibility. Complete quantitative details are in the paper on arXiv (2607.17675, July 19, 2026). The base model is Wan2.2-T2V-A14B, 14 billion parameters.

Limitations. The paper is fresh and the 30+ demos available are curated by the authors themselves: independent evaluation on real-world use cases is missing. The system is tested on a single base model, and generalization to other video diffusion architectures remains to be shown. Inference requires significant hardware, which limits access to those with datacenter-class GPU resources. For anyone producing video with AI today, ShotPlan is a signal of where research is heading, and a tool in development for your workflow tomorrow morning.