[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-0599b775-ac17-49d2-aebd-a16f531c7168":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"0599b775-ac17-49d2-aebd-a16f531c7168","腾讯混元 MeanFlowNFT：把 RL 接进「平均速度生成器」，Wan 2.1 4 步反超 50 步 LongCat-Video RL","短视频模型的开销正在被「少步生成」路线重塑。DiffusionNFT 用前向过程的负似然估计做 RL 后训练,避开了反向轨迹和似然估计的高昂成本,但它原本针对的是「瞬时速度」的扩散模型;而 MeanFlow 这种「平均速度」生成器走的是另一条路——靠预测一段时间区间上的平均速度,把采样步数压到 4 步甚至 1 步。两套公式不同,RL 优化目标一直接不上。腾讯混元 7 月 16 日公开的 MeanFlowNFT(arXiv 2607.15273)把这个缺口补上了。团队用 MeanFlow identity 在「平均速度」和「瞬时速度」之间建了一座桥:训练时构造一个 induced 瞬时速度预测器,把 DiffusionNFT 的目标函数套上去;采样阶段仍然只跑平均速度,所以原本的少步生成优势一个字符不掉。论文还证明 MeanFlowNFT 继承了 DiffusionNFT 的策略单调改进保证。数字最有说服力:Wan 2.1 视频模型上,4 步 MeanFlowNFT 把 VBench 推到 84.33,直接压过 50 步的 LongCat-Video RL(82.57);SD3.5-M 图像生成上,8 个评估维度里 6 个超过此前所有 RL 调优的少步生成器。项目已开源(github.com\u002FHarahan\u002FMeanFlowNFT)。把 RL 后训练的成本曲线从「必须配合多步 ODE solver」改成「少步也能稳定对齐奖励」,对实时视频生成、交互式视频 Agent 这些场景是直接利好——这篇工作值得 Hunyuan 之外的从业者留意:它不是单一模型的升级,而是给整个少步生成路线补上 RL 这一课。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.15273","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation","2026-07-16T12:00:00Z","2026-07-19T00:12:29.890817Z","2026-07-19T00:12:29.890838Z",true,"agent",6]