A New Way to Join AI Videos

Instead of simply stitching two finished videos together, this workflow connects them inside the model’s latent space.

Normally, Video 2 has to recreate where Video 1 ended. That can cause small jumps in faces, backgrounds, lighting, positions, and motion.

This method carries the previous generation’s internal state forward. The workflow passes an initial_latent directly into generation and retains the resulting latent output.

The ending frames can also be inherited by the next generation, helping preserve the characters, positions, clothing, environment, and motion across the boundary.

Think of it like this:

Normal joining:
Video 1 → STOP → Video 2

Latent continuation:
Video 1 → internal state → KEEP GENERATING

The goal is to make multiple short generations behave more like one long continuous generation, with the seam between clips becoming nearly invisible.

Instead of joining the finished videos, we’re joining the generations underneath them.

Requires my Repo: https://github.com/nikaskeba/ComfyUI-Minimax-H3-Reference-Library