Beach

Beach is a 37-second vertical clip in black and white. A fisheye camera skims the wet sand of a Baltic beach while striped windbreaks tear loose and fly off into the sky. The sea was meant to stand upright like a wall. At the end, a woman in a black 1930s swimsuit and a white swim cap dances at its foot. The soundtrack is dark instrumental techno made with YuE2, a music model running locally.

How it was made

It was a playful test of new models, about 90 minutes long. I showed the Claude Code agent a short clip from the web: a black-and-white camera flight through a gale where physics breaks down. The agent pitched several worlds; I picked the beach and the models to test, YuE2 included.

• The agent generated four versions of the track in YuE2 and kept the one with the steadiest tempo, 132 BPM.
• In MiniMax-H3-Image it turned a long set description into a start frame, the still image the first shot grows from.
• The shots come from MiniMax H3, a video model that generates picture and sound together. On the same prompts, the agent pitted our default build, nicknamed Pika, against Flash, a slimmer, faster H3. The flight is three 10-second clips, with the track written into the latent, the compressed internal form the model works on.
• Pika caught the beat better, I thought, and Flash looked more cinematic, so I suggested adding the Pika H3 Flex LoRA (a small add-on that changes how a model behaves) to Flash. Of the dance takes, I chose Flash with the add-on. The agent re-rendered the flight with it, then added the dance and the track.

What it shows

• Chain clips in the latent. At first the agent fed each new clip finished frames from the previous one: 39 to Pika, one to Flash. Flash restarted the motion almost from scratch and ignored 39 frames when given them. I asked about chaining through the latent; after the switch, clips flowed into each other on both models.
• Rhythm is partly method. To my ear, plain Flash kept the beat in the dance, so the agent concluded the flight’s weaker sync probably came from how the clips were joined. A take with a different sampling setting looked best but lost the beat and was dropped.
• Not everything worked. The start frame shows the water as a vertical wall, but during the flight both H3 variants soon bent it into an ordinary fisheye horizon. The LoRA version also lost the beach chairs from the second shot and was slower to render: about 143 seconds per clip instead of 110–118.