Local AI video: 480p in 90 seconds

I tested every optimization the internet recommends. SageAttention v2, TeaCache, torch.compile, GGUF quantization.

The win came from MagCache, out of a NeurIPS 2025 paper. It detects when a transformer block’s activations have not meaningfully changed and skips recomputing them.

Combined with 4-step distillation, MoE-aware GPU splitting and block-level distribution, Wan 2.2 14B went from 75 minutes to 1:26 at 480p. Two used 3090s. No cloud.

Is it as good as the best closed-source models on better hardware? No. It does show how many optimizations are still sitting there unfound.