Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8
Reported by Hacker News1 source ↓
Video Rebirth presented HyperFlow H3 Flash, a compressed version of MiniMax-H3 that generates video with synchronized stereo audio. The project describes a 14B-parameter, six-step INT8 model derived from a 31B model.
Key points
- HyperFlow H3 Flash generates video and synchronized stereo audio
- The compressed model is 14B parameters and runs in six steps
- The project recommends GGUF for PCs with 64 GB of system RAM or less
- INT8 convrot is recommended for data-center GPUs or 24 GB-plus RTX GPUs with at least 69 GiB of system RAM
Sources
Details
Hacker News's Show HN post says HyperFlow H3 Flash modifies the MiniMax-H3 transformer to create a smaller, few-step version. The post links to model weights, ComfyUI nodes, a GGUF setup guide and an eight-step LoRA for full-size MiniMax-H3.
The project's benchmark script measures each pipeline stage and peak memory per GPU. Reported runs include text encoding, transformer processing, and video and audio VAE decoding at 1344×768 resolution and six steps, for clips lasting 5.2 or 15 seconds.
Written from coverage as of
Images
- Hacker News