LightX2V Enables 14B Video Model on Single RTX 5090 with 720p Real-time Generation
新智元 · wechat · 2026-08-07
The LightX2V team introduced an extreme acceleration framework for the Wan2.2-A14B video generation model, drastically reducing computational load and VRAM peak usage via step distillation, NVFP4 quantization-aware training, and dynamic sparse attention.
- Core Optimizations: Uses PhasedDMD to compress denoising steps from 40 to 4. Combines NVFP4 quantization and dynamic sparse attention to lower per-step compute overhead.
- Single GPU Performance: The 14B model runs efficiently on a single RTX 5090, generating a 5-second 720p video in just 22.5 seconds, achieving over 100x speedup.
- Multi-GPU Real-time Generation: On an 8-GPU setup, T2V and I2V for 5-second 720p videos take only 3.8s and 4.5s respectively, both shorter than playback time, achieving up to 705.8x acceleration.
More from Infra
- Amazon Proposes Per-Request Validated Evidence Caching for Production RAG — _reachsumit · 2026-10-16
- LinkedIn's TransX report: sequence transduction for recsys, +6.02% CTR in A/B tests — _reachsumit · 2026-10-16
- llama.cpp PR boosts CPU inference speed with merged AVX2 tiled accumulators — jacek2023 · 2026-10-16
- Zero's next-gen IVM engine runs in a single DO, live-querying a 100GiB R2 database — threepointone · 2026-10-16
- Meme: driving to the chip factory in 2028 captures AI's compute anxiety — beffjezos · 2026-10-16
- One volunteer maintains the time zone file behind 4 billion phones — danbri · 2026-10-16