Research method for multi-shot text-to-video with frame-accurate cuts, fades and camera moves, built on Wan2.1-T2V-14B.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
ShotPlan is a research method for multi-shot video generation, built on the Wan2.1-T2V-14B text-to-video backbone. It adds a small number of learnable "planning tokens" to the diffusion transformer's input sequence so that the timing of events inside a clip can be specified per frame. Each event type (hard cut, fade-in begin, fade-in end, camera movement) has one learnable embedding, which is replicated once per requested event and attends jointly with the video tokens.
The tokens use fractional temporal positions (called FRoPE), which lets them point between latent frames despite the 3D VAE's 4× temporal compression. They are removed after the final DiT block, so the output video has the same shape as the base model's. Users supply a single prompt with per-shot captions plus timestamps for cuts, cross-fades or camera moves (e.g. circle-left, zoom, dolly, truck).
On the authors' multi-shot benchmark, ShotPlan reports a transition deviation of 0.64 frames, compared with 1.12 for MultiShotMaster and 2.71 for HoloCine, and also reports gains in text alignment, scene consistency and narrative coherence. Demo videos are 832×480 at 81 frames. A YouTube creator summary mentions Wan 2.1 and 2.2 variants and released training code, but no repository or weights links appear in the provided material, so these claims could not be verified here.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source multi-shot video generator that follows precisely timed cuts, fades and camera movements specified per shot. Consistent characters; variants for Wan 2.1 and 2.2; training code released.
Closest neighbours by meaning, not just by category.
Research
Research method for long-context video generation using hierarchical latents and coarse-to-fine rollout.
Model · Qualcomm AI Research
Qualcomm research model that runs a distilled, pruned Wan2.2 5B video diffusion model on a mobile device.
Model
Hybrid softmax and linear attention variant of MiniMax H3 for faster long video generation, from UC Berkeley and Impossible, Inc.
Library · Tencent
Positional encoding method that gives video diffusion transformers precise camera control from an image and camera path.
Model
Research project for real-time autoregressive generation of multi-shot video narratives from shot-level prompts.
Model · ByteDance
ByteDance unified video generation and editing framework pairing an MLLM semantic planner with a DiT renderer.
