2 ms·
Show HN: The Fastest Local H3 Video Generation on Mac
960×544, 124 frames, 6 DiT steps with Turbo LoRA:
On M5 Pro MacBook Pro 16", 24 GB RAM:
Vpipe 5 min 0 sec vs. H3.c 7 min 19 sec.
For comparison, H3.c is probably the best-known “hardcore” H3 implementation out there.
Vpipe runs H3 through our own C++/Metal inference stack — no Python or third-party tensor runtime in the forward pass. We use weight streaming with resident blocks to fit the 33B model on a 16GB Mac, with M5-specific acceleration through NAX matmul2d.
If you know of a faster H3 implementation on Apple Silicon, we’d love to benchmark against it.
- t-go 3mo agofor quality comparison, check out this: https://www.youtube.com/shorts/i3F_Fzgt1UM https://www.youtube.com/shorts/i3F_Fzgt1UM