This workflow demonstrates a trick to add reference image anchor (e.g. first image) when generating long video through looping with Minimax H3 FL2VA model. (i.e. extend video from the last few frames of the previous video)

Unlike Wan2.2 SVI, H3 does not offer adding reference images to FL2VA model. When a video is generated from the last portion of a previous video, the first portion is missed. This leads to e.g. change of character appearance.

Here is the idea of the trick:

1. Replace the first image(s) of "the last portion of video 1" with your reference image(s) e.g. first image of video 1. Video 2 will begin with 1 reference image + remaining last portion of video 1.

2. In video 2 prompt, divide it into 2 shots. [Shot 1] contains only 1 frame of reference image. [Shot 2] At 0.05 seconds, describe the main plot.

Since Minimax model has already seen your reference image (e.g. character) in [Shot 1], it will follow the guidance instead of guessing blind.

In theory, you should be able to add multiple reference images. Reference images and last protion from previous video will be trimmed in the end.

Cheers.