Lightx2v's NVFP4 checkpoint speeds up WAN 2.2 14B by 50x, but quality drops

Lightx2v has released a new NVFP4-Sparse checkpoint for the WAN 2.2 14B video generation model, claiming dramatic inference speedups that could revolutionize rapid prototyping. According to their benchmarks, the NVFP4 variant reduces generation time from 734 seconds to just 14.15 seconds at 480p (51.9x faster) and from 2668 seconds to 45 seconds at 720p (59.3x faster). These figures suggest near-real-time video generation for low-resolution outputs, a leap forward for iterative design and content creation.

However, early community feedback indicates a clear trade-off: motion quality in NVFP4 outputs is noticeably inferior to the full-precision version. The original poster noted that the motion coherence and fluidity are significantly degraded, likely due to aggressive quantization and sparsification. Despite this, the speed improvements are so stark that many users will find the trade-off acceptable for early drafts or storyboarding. Integration into ComfyUI is anticipated soon, which would lower the barrier for experimentation. The key question remains whether the speedup numbers include the Lightning variant or are for the standard model, as Lightx2v hasn't clarified this detail.