VibeusA 15-second generated travel montage, moving from a busy street to... - Vibeus
A 15-second generated travel montage, moving from a busy street to a food stall, a shrine gate and a night-time overlook, made locally on a Mac. TgoAI’s Valtz demo combines Pika H3 Flex with NVIDIA’s Open-H3-Regenerate-2K: generate a 768p draft, then refine it at higher resolution. The creator reports 12 minutes 19 seconds on an M5 Ultra for generation and regeneration together, with a final output of 1920 × 1088. The exact memory configuration is not given.
Pika H3 Flex distils MiniMax H3’s 50-step generation into six steps. It is released as two LoRA adapters for the H3 base, covering first/last-frame and reference-driven generation, rather than a separate architecture. The published sampling schedule matters: simply setting an ordinary H3 sampler to six steps is not equivalent. VPIPE reads the adapter’s manifest to choose that schedule.
Open-H3-Regenerate-2K then encodes the draft, enlarges its latent directly and uses three diffusion steps to redraw detail at the target size. It avoids the decode–resize–encode round trip of pixel-space upscaling. VPIPE’s regeneration pipeline carries the draft’s soundtrack into the refined video. The complete original Reddit clip, including its audio, is attached.
Valtz is an open-source native macOS creation app powered by VPIPE’s Metal inference engine. Its README requires Apple Silicon and macOS 26 or later. Valtz code uses MPL-2.0 and VPIPE Apache-2.0; model weights are downloaded separately under their publishers’ terms, including H3’s territorial and commercial-use conditions. The demo is one creator’s timed run, not a general Mac speed comparison. The discussion also raises face-consistency problems in wide shots; this mostly close-range sample does not resolve them.