Alibaba open-sources 14B video model that bridges start and end frames

Wan2.1-FLF2V-14B is Alibaba's open-source 14B-parameter video generation model, built around a very straightforward trick: hand it a starting frame and an ending frame, and it will generate the transition video in between.

What it can do

Standard text-to-video and image-to-video tools often feel like gambling. You know where the shot starts, but having it land precisely on a specific closing visual usually takes endless rerolls.

Wan2.1-FLF2V fixes this by locking down both bookends. Creators can define the exact first and last frames, and the model fills in the action, giving short-video makers predictable control over shots and scene bridges.

On the official Wan website, users can prompt the model to generate a 5-second video at 720p resolution for free. The model weights and source code are already up on GitHub, Hugging Face, and Alibaba's ModelScope hub.

How it works

Bridging two distinct static images without jarring glitches or mutated limbs is notoriously tricky in generative video.

Wan2.1-FLF2V pulls this off by integrating an additional control adjustment mechanism. It treats the user-supplied start and end frames as strict conditioning inputs and injects their semantic features throughout the generation pipeline.

By keeping tabs on both endpoints at every step, the model preserves style, subject structure, and scene context while resolving complex character and camera movements across the timeline.

This release builds on Alibaba's Wan2.1 foundation family. The company open-sourced four Wan2.1 models back in February 2025, racking up over 2.2 million downloads across Hugging Face and ModelScope. The series was the first to support both Chinese and English text effects and currently sits at the top of the VBench leaderboard.

Why it matters

For animators, digital artists, and short-video editors, this is an immediate workflow booster. Instead of manually stitching cuts or rolling dice on generative continuations, you can treat AI as an automated inbetweening tool between two storyboard keyframes.

While many open video models are still stuck competing on generic image-to-video motion, Alibaba made a smart play by releasing a dedicated 14B checkpoint built specifically for start-and-end frame control. Running a 14B model locally still requires hefty VRAM, and output is currently capped at 5 seconds at 720p, but it offers a far more controllable canvas than most open alternatives today.