AI 精选动态
智能评分 60
FastWan-QAD:单卡RTX 5090生成5秒视频仅1.8秒
AI 推荐理由
该结果验证了量化感知蒸馏在消费级硬件上的高效性,值得关注其开源实现和性能复现。核心解读
Hao AI Lab发布FastWan-QAD视频生成模型系列,采用FastVideo的量化感知蒸馏(QAD)技术,在单张NVIDIA GeForce RTX 5090上端到端生成5秒480P视频仅需1.8秒,并同时提供针对RTX 4090优化的FP8检查点。
全文
(1/5) 5 seconds of video. 1.8s seconds of generation. One NVIDIA GeForce RTX 5090 on FastVideo. 🤯🚀
- FastWan-QAD, a new family of video generation models
- Trained with FastVideo's Quantization-Aware Distillation (QAD) recipe.
- Powered by FastVideo, we push a single NVIDIA GeForce RTX 5090 to its absolute limit: generating a 5-second 480P video in 1.8s end-to-end!
📜 Blog: https://t.co/xs7fPvbGuh
💻 Code: https://t.co/5YnBSko00N
💽 Model: https://t.co/NNbl9Og6bw
https://video.twimg.com/amplify_video/2069491796284194816/vid/avc1/832x480/HAP6h_UIK9SVBw5j.mp4?tag=28
Hao AI Lab (@haoailab): (4/5) All these optimizations mean generation is lightning fast. ⚡
Clocking in at 1.8s for a 5-second video on a single RTX 5090, proving that ultra-low latency video generation is entirely viable on consumer hardware. Broad hardware support is a priority for us, so in addition to NVFP4 checkpoints for the RTX 5090, we are also releasing a dedicated FP8 checkpoint fully optimized for RTX 4090 users.
Hao AI Lab (@haoailab): (5/5) We are driven by a belief that video generation should be fast and fully accessible locally.
All code for inference is available in the FastVideo repo! Check out our blog post for step-by-step instructions on how to run it: https://t.co/xs7fPvbGuh
Huge thanks to the FastVideo team, @NVIDIAAI , and @mbzuai for helping push open-source video generation forward!