Frame In-N-Out

Frame In-N-Out expands the first-frame to a larger canvas, where it allows users to assign motion trajectories to existing objects and introduce new identities that enter the scene with specified trajectories.
The model we used here is Wan2.2-5B V1.6 trained on our Frame In-N-Out control mechanism.

Easiest way: Choose one from Examples below and then simply click Generate.

❗️❗️❗️Instruction Steps:
1️⃣ Upload your Input Image 🖼️ . Next, set Resized Height for Input Image and Resized Width for Input Image for the size you want.
2️⃣ Set Top-Left Expand Height, Top-Left Expand Width, Bottom-Right Expand Height, and Bottom-Right Expand Width for the expansion amount.
      The Canvas Height (Resized Height + Top-Left Expand Height + Bottom-Right Expand Height) and Canvas Width (Resized Width + Top-Left Expand Width + Bottom-Right Expand Width) should be the multiplier of 32.
      Recommend Canvas Height = 704 and Canvas Width = 1280 for the best performance (pre-trained model default resolution).
3️⃣ Click Build the Canvas.
4️⃣ Provide the motion trajectory of the object by clicking on the Expanded Canvas 🖼️ . You can make additional trajectory for the same object by clicking Add New Traj Line (Same Obj). Reset by Clear All Traj.
5️⃣ Provide the Identity Reference image and its trajectory (optional). Since image is segmented by SAM first (providng center point as query), it will be nice for the inputs to be center cropped.
      New instance trajectory can be done by clicking Add New Instance (New Obj, including new ID).
6️⃣ Write a detailed text prompt.
7️⃣ Click the Generate! button to start the Video Generation.

If Frame In-N-Out is helpful, please help star the GitHub Repo. Thanks!