[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-virdm-teacher-free-video-post-training":3,"topics-all":38,"news-related-989509cd-0977-4ddf-9443-8ef49485befd":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"989509cd-0977-4ddf-9443-8ef49485befd","ViRDM:三网络视频蒸馏,砍到只剩一个生成器","arXiv 新论文 ViRDM 把少步因果视频生成的后训练从三网络蒸馏改成只训生成器:不再需要冻结教师和在线批评家,直接对预计算的表示分布做匹配。论文报告官方 VBench 总分 84.87,比此前最佳少步因果基线高 0.36,全程只花 20 次生成器更新、16 个 A100 卡时。","少步自回归视频扩散是目前流式生成的主流路线:视频被拆成按时间排序的块,块内做多步去噪,块间复用 KV 缓存,天然支持边生成边播放。但要把它从「能用」压到「能实时」,主流做法是后训练蒸馏——而这条路的开销一直降不下来。\n\n## 三网络蒸馏到底贵在哪\n\nSelf Forcing、Causal Forcing 这类基于 DMD(分布匹配蒸馏)的方案,每次更新要同时维护三个视频扩散网络。论文里给了一笔账:蒸馏一个 Wan 2.1 1.3B 的生成器,典型配置还要额外挂一个冻结的 14B 教师和一个在线更新的 1.3B 批评家。教师给出目标分布的得分,批评家拟合学生分布的得分,两者相减才是训练信号——三个网络的维护与前向,让每次更新都在算力和显存上付出成倍代价。\n\n## ViRDM 的思路:对齐预计算的表示分布\n\nNortheastern University 与 Adobe Research 的团队(第一作者 Zichong Meng,实习期工作)提出的 ViRDM 换了个问法:能不能不要教师、不要批评家,只让生成器去对齐一份提前算好的目标分布?思路来自一步图像生成里的 RDM(表示分布匹配),但直接搬到视频上会撞三堵墙:端到端梯度路径显存装不下、图像那套优化配方在视频上失效、表示分布管不住时序动态。\n\nViRDM 的配方逐堵拆墙:用随机截断的干净出口,让每次更新只回传一个被选中去噪步的干净端点预测;换轻量 VAE 解码器;用分阶段向量-雅可比积把生成器、解码器、表示编码器三段反向图错开驻留显存。论文还摸出了视频特有的优化区间:每个批次 8 到 64 条新生成视频就够用(图像 RDM 报告的最优区间在 2048 条以上),且初始化必须来自因果适配过的模型——表示分布匹配能把现成的因果生成器打磨到位,但没法把双向模型凭空改造成因果模型。针对动态维度偏弱的问题,再补一个轻量动力学正则。\n\n## 数字层面:20 次更新,16 个 A100 卡时\n\n论文报告的官方 VBench 总分 84.87,比此前最佳的少步因果基线 Causal Forcing 高 0.36,Dynamic Degree 达到 72.02;若按不含动态维度的口径计算,Total 为 85.77,也超过 Self Forcing 此前的 85.34。训练成本一侧:8 张 A100 跑 2 小时合计 16 卡时,共 20 次生成器更新,每次约 6 分钟,每卡峰值显存 48.3 GB;换单张 80GB A100 靠梯度累积也能跑,峰值 68.5 GB。同一配方还能外推:后续块一步去噪的因果生成拿到 84.27 的同档最佳 Total,双向生成 1\u002F2\u002F4 步分别为 83.12\u002F84.53\u002F84.56。\n\n## 冷静看的部分\n\n论文自己的 Limitations 写得很直白:这是后训练配方,不能替代因果预训练;主实验只验证到 Wan2.1-1.3B 主干和 81 帧视频,更大规模、更高分辨率、更长时序还没测;方法还继承冻结表示编码器和固定参考分布的归纳偏置。另外「超过基线 0.36 分」是作者在自家实验设置下的自报对比,第三方复现结果需要等时间检验。\n\n对做视频生成工程的人来说,真正的信号是:蒸馏不再是「三网络堆算力」的唯一解,表示分布匹配把少步因果后训练的入场门槛从集群级压到了单卡可试。下一步值得盯的是这套配方在更大主干上的复现——如果 16 卡时的量级能保持,少步因果视频生成的成本曲线会再往下压一截。\n\n参考:arXiv:2609.28923(https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.28923),项目页 neu-vi.github.io\u002FViRDM\u002F","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.28923","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2e8820ca-3be4-4298-b54f-62548b774b40","en","ViRDM: Few-Step Video Post-Training Without Teacher or Critic","ViRDM drops the teacher and critic from few-step video post-training: 84.87 VBench Total, 0.36 over the prior causal best, 20 updates, 16 A100 GPU-hours.","Few-step autoregressive video diffusion is the mainstream route to streaming generation: a video is factorized into ordered temporal blocks, diffusion runs inside each block, and completed blocks are cached and reused to condition future generation. That causal, streaming interface is what interactive applications want. But pushing it from \"works\" to \"real-time\" has mostly meant post-training distillation, and the cost of that route has stayed stubbornly high.\n\n## Why three-network distillation is expensive\n\nDMD-based methods such as Self Forcing and Causal Forcing maintain three video diffusion networks during each training update. The paper lays out the bill: distilling a Wan 2.1 1.3B generator typically requires an additional frozen 14B teacher and an online 1.3B critic. The teacher supplies the target-distribution score, the critic fits the student-distribution score, and only their difference yields a training signal. Maintaining and evaluating those networks makes every update expensive in both compute and memory.\n\n## The ViRDM idea: match a precomputed distribution\n\nViRDM, from Northeastern University and Adobe Research (first author Zichong Meng, done during an internship at Adobe Research), flips the question: can we drop the teacher and the critic entirely, and post-train only the generator against a precomputed target distribution? The idea transfers representation distribution matching (RDM) from one-step image generation, but a literal transfer hits three walls: a memory-intractable end-to-end gradient path, an image-derived optimization recipe that fails on video, and representation distributions that underconstrain temporal dynamics.\n\nThe recipe dismantles the walls one by one. Stochastically truncated clean exits select one denoising step per update, so only that step's clean endpoint prediction is backpropagated. A lightweight VAE decoder and staged vector-Jacobian products keep the generator, decoder, and representation encoder backward graphs from coexisting in memory. The team also maps out a video-specific regime: fresh batches of 8 to 64 generated videos already suffice, versus the broad optimum above 2,048 samples reported for image RDM. Initialization must come from a causally adapted model, representation distribution matching can refine an existing causal generator but cannot convert a bidirectional checkpoint into a causal one. A lightweight dynamics regularization closes the remaining gap on motion.\n\n## The numbers: 20 updates, 16 A100 GPU-hours\n\nThe paper reports 84.87 Total on the official VBench evaluation, 0.36 above the previous best few-step causal baseline, Causal Forcing, with Dynamic Degree at 72.02. Without the Dynamic Degree dimension, the model reaches 85.77 Total, surpassing the 85.34 previously achieved by Self Forcing. On the cost side: two hours on eight A100 GPUs (16 GPU-hours), 20 generator updates at roughly six minutes each, peak memory of 48.3 GB per GPU, and a single 80 GB A100 configuration with gradient accumulation peaking at 68.5 GB. The same recipe extends further: with one denoising step for subsequent blocks it obtains the best Total of 84.27 in that setting, and on bidirectional generation with one, two, and four steps it scores 83.12\u002F84.53\u002F84.56.\n\n## Where to stay skeptical\n\nThe paper's own limitations section is blunt: this is a post-training recipe, not a replacement for causal pretraining; primary experiments cover only the Wan2.1-1.3B backbone and 81-frame videos, leaving larger scales, higher resolutions, and longer horizons untested; and the method inherits the inductive biases of its frozen representation encoder and fixed reference population. The \"+0.36 over the baseline\" figure is the authors' own comparison under their experimental setup, so third-party replication is still pending.\n\nFor video-generation engineers, the real signal is this: distillation is no longer the only game that requires stacking three networks on a cluster. Representation distribution matching pushes the entry ticket for few-step causal post-training down to single-GPU territory. The next thing to watch is replication on larger backbones — if the 16-GPU-hour order of magnitude holds, the cost curve of few-step causal video generation gets pushed down another notch.\n\nReference: arXiv:2609.28923 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.28923), project page neu-vi.github.io\u002FViRDM\u002F","virdm-teacher-free-video-post-training","2026-09-26T19:06:45Z","2026-09-26T19:06:48.222391Z","2026-09-26T19:06:48.222407Z",true,"agent",624,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"4349edac-997e-4980-95a0-58054a28a9dd","视频生成投机去噪:1.7 步顶 4 步,提速 1.5 倍","speculative-forcing-video-denoising","2026-10-16T19:11:48+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"0b0af156-7b5c-47d8-bfc2-de8cc0b22ce2","混元开源Prism:稀疏注意力让2K音视频训练提速2.5倍","prism-2k-video-audio-sparse-attention","2026-10-13T23:13:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"05cb84fb-1408-4130-bfe9-9d924461c549","DiffusionART:扩散模型RL微调省12倍GPU时","diffusionart-reward-tilting-gpu-efficiency","2026-10-12T13:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"24d56562-4886-4169-8aa9-773f5fa34e2f","GRACE 把视频潜空间压 8 倍:提速 11 倍,跑分反超","grace-latent-compression-video-diffusion","2026-10-08T15:10:59+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a2e06c5f-ab65-468f-b936-c654ff75aa38","14 秒视频 9 秒出片:Video DeltaNet 给 H3 换混合注意力","video-deltanet-h3-hybrid-attention","2026-09-18T23:08:51+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"caed836e-2168-418f-b5c1-bde3ce962e66","Mask Forcing 往蒸馏 rollout 里掺干净 token:修视频生成的模式坍缩,指令遵循最高涨 6.5 分","mask-forcing-video-diffusion-distillation","2026-09-09T23:08:37+00:00"]