H3 Long Video Pixaroma

When to use it

Only when your MiniMax H3 video is longer than about 15 seconds, for example a whole song or a long talk with H3 Audio Sync.

For a video of 15 seconds or less you do not need it. If it is on the wire anyway, nothing changes: the video is made in one piece as usual.

What it does

MiniMax H3 was trained on clips of up to about 15 seconds. Longer videos used to mean making several clips and joining them, and every join showed: the next clip came out a little darker and softer, which you see as a flash and a drop in quality.

This node makes the whole video as ONE piece instead. On every step the model looks at it in overlapping windows of about 10 seconds, each with its own part of your song, and the windows are blended where they overlap. Because every window is worked on at every step, they agree with each other all the way through, so there is nothing to join.

It is also lighter and faster than one long run. A minute at 480x864 took 4.5 minutes instead of 9.4 on a 4090, and with 12 GB of graphics memory free it ran at the same speed, while one long run of a minute could not even finish its first step.

Where it goes

On the MODEL wire, after your LoRAs and BEFORE H3 Audio Sync: model, then H3 Long Video, then H3 Audio Sync, then the sampler. In that order Audio Sync can see the windows and skips its 15 second warning.

Then simply make the video as long as you want: set the length on the latent node (or Duration Pixaroma) to the whole song, and cut the same length of song with Load Audio Pixaroma. Everything else in the workflow stays the same.

Example

Open the EP41 workflow "Minimax H3 - Audio Sync - Make a Picture Sing". Add H3 Long Video Pixaroma between ModelAttentionBackend and H3 Audio Sync. In Duration Pixaroma open the gear and add 30 to the lengths, then click 30. In Load Audio Pixaroma pick a song that is at least 30 seconds long. Run: you get one 30 second singing video with no joins.

Window and overlap

Window is how much of the video the model looks at at once. 10 seconds is the tested default and it held up best; 5 second windows worked too but drifted a little more in brightness. Overlap is how much two neighbouring windows share; 1.5 seconds is the tested default.

Both are snapped to H3's frame grid, so 10 becomes 10.12 and 1.5 becomes 1.62. A video no longer than one window is made in one piece exactly as without this node.

Tips

• Use MiniMax H3 Reference to Video. Your picture reaches every window. With Image to Video's first frame, only the first window sees the picture.
• Use your real song or voice with H3 Audio Sync. Every window sings its own part of the track, so the sound has no seam at all.
• Each window leans a little toward the pose in your picture. A striking pose there, like a hand on the chest, can come back now and then in a long video (more at 480x864 than at 576x1024). A picture in a relaxed pose avoids it.
• Make sure your song is as long as the video. Where the track ends, H3 Audio Sync fills the rest with silence, and the character stops singing there.
• It makes new video from scratch. Running it over an upscaled low-resolution video does NOT sharpen it: H3 keeps a soft input soft.