[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-specslot-pixel-slot-global-attention-decoupling":3,"topics-all":38,"news-related-c03ecfed-2b3f-4831-bbc0-fad67a993917":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},"c03ecfed-2b3f-4831-bbc0-fad67a993917","SpecSlot:像素生成不必让细节都走全局注意力","像素空间生成模型 SpecSlot 把图像词元划分方式改写成预算分配:锚点密度与频谱细分分开算,每个锚点只有最低频那一槽进入全局注意力,其余槽都留在锚点内部做局部交互。论文自测在 ImageNet-512 上 FID 9.14,比同算力对照低 2.84 分且算力更省,文生图 DPG 基准分高出对照 1.06 分。","像素空间生成模型绕开了单独训练的自动编码器，直接从噪声学像素，代价是整张分辨率的图像都得表示成 Transformer token。大多数这类模型一个图像块占一个 token，于是同一张网格同时决定了模型读什么、每个 token 预测什么、以及多少个 token 要挤进全局注意力——想看更细，注意力代价就直接跟着涨。256×256 的图上，patch 从 16×16 降到 8×8，token 数从 256 涨到 1024，稠密注意力的二次项就放大 16 倍。\n\n问题出在这种耦合上：token 在全局注意力里能跟谁通信，是由它坐在网格的哪个位置决定的，而不是由它最后产出什么决定的。\n\n## 把切分方式改写成预算分配\n\n北京大学、南洋理工大学和商汤的团队提出 SpecSlot，换的正是切分方式。它把图像 token 总数写成 N = W × K：W 是空间锚点密度，K 是锚点内部的频谱细分。每个锚点展开成 K 个槽，每个槽固定负责一段互不重叠的 DCT 系数带；输出侧再叠一个空间残差做局部修正。生成状态、目标和损失都仍然留在像素空间。\n\n通信按角色来分。每个锚点里最低频的那个槽是载体，只有它进入全局注意力，其余槽只在锚点内部局部交互。调度顺序是全局、全局、局部、局部，重复三轮，一共六个全局块加六个局部块，维度一致。于是全局序列长度只跟 W 走，不再跟 W×K 走：ImageNet-256、s=16、K=4 的配置下，全局序列是 W+C = 288 个 token，而同条件下把所有槽都放进全局的方案需要 1056 个。对应的二次项从 12(WK+C)²d 变成 6(W+C)²d + 6WK²d。\n\n这一层被论文概括为\"tokenization as allocation\"。它同时松开了两件此前绑在一起的事：读输入的密度由输入网格决定，输出角色由锚点内的频带决定，谁也不再绑架谁。\n\n## 优势要到 512 分辨率才真正拉开\n\n论文自己的评测里，ImageNet-256 上 SpecSlot W256\u002FK4 拿到 6.59 FID、129.20 GFLOPs，只比同算力对照 JiT-CM 的 6.68 好 0.09 分；作者说这一差距落在五个随机种子的 95% 区间内，基本算持平。真正的质量赢家仍是更贵的 JiT-B\u002F8：4.34 分、218.50 GFLOPs。\n\n到了 ImageNet-512，差距才打开：W256\u002FK4 是 9.14 FID、135.64 GFLOPs，比同算力对照的 11.98 低 2.84 分，前向 FLOPs 还少 3.9%；跟更贵的 JiT-B\u002F16（8.45 分、219.56 GFLOPs）之间只差 0.69 分，算力便宜 38.2%，吞吐 12.69 对 7.93 img\u002Fs。作者给的解释是：512 下每个锚点覆盖的像素是 256 的四倍，而全局序列长度始终固定在 288，多出来的细节被留在了锚点内部。\n\n文生图那边，256 分辨率下 DPG-Bench 62.21，比 JiT-B\u002F8 的 61.15 高 1.06 分，前向算力从 218.77 降到 129.86 GFLOPs，少 40.6%。但同一个模型在 GenEval 上是 26.99，低于 JiT-B\u002F8 的 29.78。同一套架构换文本条件就能用，不需要针对文生图改结构，代价是在另一条对齐指标上让了分。\n\n## 一个容易被跳过的细节：算力对齐不等于速度对齐\n\n路由对比表最能说明这篇工作的性质。同一 W256\u002FK4 配置下，GGLL 是 6.59 FID、129.20 GFLOPs、9.06 img\u002Fs；算力匹配的稠密全局对照 Dense-global-CM 是 6.81 分、128.96 GFLOPs，却只有 5.80 img\u002Fs。两者实测前向 GFLOPs 几乎一样，墙钟吞吐却差了一大截——在 20、30、50 个 Heun 区间三个预算上，GGLL 的吞吐都是对照的 1.47 到 1.49 倍。\n\n再往上一层，把全部 1024 个槽都放进全局注意力的 Full-global 拿到 5.84 分，是这张表里最好的 FID，代价是 225.82 GFLOPs，比 GGLL 高 74.8%。也就是说，SpecSlot 赢的不是\"同算力下的最高分\"，而是把\"全都全局通信\"这条路的成本砍掉 42.8% 之后，仍然守住了比算力匹配对照更好的质量。它选的是中间那个受注意力约束的工作点，不是最左边的那个。\n\n## 机制自证：低频确实先被恢复\n\n论文给了一套机制证据。固定 W256\u002FK4\u002FB1536，只改哪个槽拥有跨锚点访问权：槽 0 得 6.59，换成槽 1、2、3 分别是 6.91、7.21、6.97，全都更差。输出分支的探针显示，频谱分支自能量更大，而空间分支的份额随流过程从 0.07 涨到 0.23；在 t=0.90 处拿掉任一分支，速度 MSE 分别涨 32.1 倍和 64.1 倍，在 t=0.05 处则是 1.1 倍和 2.1 倍——谱骨架从头到尾都在，空间修正是后期才补上来的高频。\n\n分配消融则显示，固定槽预算 2048 时，锚点越多分越低（W64\u002FK32 的 7.28、W256\u002FK8 的 6.28、W1024\u002FK2 的 5.66）；而锚点固定为 256 时，把 K 从 2 加到 4 再加到 8，增益一次比一次小，最后一次翻倍只换来 0.31 分。\n\n## 我的判断\n\n这批结果目前只有一个来源：论文、官方代码库和项目页同属一个团队，所有分数都是作者自测，还没有第三方复现；论文本身标注在 ICLR 2027 评审中。更合适的读法是把它当成一个方法学提案，而不是一条已被验证的结论。\n\n但它指的方向值得留意。过去两年像素空间生成的讨论基本围绕\"怎么把自动编码器彻底拿掉\"，这篇把问题换成了\"细节和通信预算怎么解耦\"。它在 256 上没能赢过更贵的 JiT-B\u002F8，赢的是在受约束区间给出不被支配的工作点；优势明显出现在 512 上，这恰好暗示解耦的收益跟分辨率一起放大，而不是均匀铺开。\n\n对做生成系统的人来说，更实用的可能是它顺手给出的那条对照结论：算力对齐的两个模型，墙钟速度可以差出一截。压 FLOPs 不等于压延迟，尤其是当你真正改的是注意力的形状而不只是它的规模。下一步要看的，是这套频带分工能不能撑到更大规模，以及在视频这类时序数据上，低频与高频的分工是否还成立。\n\n论文与代码：arXiv 摘要页 [arXiv:2610.40647](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40647)，项目页 [yangjiongfan.site\u002Fspecslot](https:\u002F\u002Fyangjiongfan.site\u002Fspecslot\u002F)，代码仓库 [JiongfanYang\u002FSpecSlot](https:\u002F\u002Fgithub.com\u002FJiongfanYang\u002FSpecSlot)。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40647","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"c883fd20-1d66-4fb7-9fc7-320fa7f87023","text-to-image",{"id":22,"name":23,"slug":23,"description":14,"color":14},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"eabda3de-e545-484f-9e1a-c9e7cdd12eb6","en","SpecSlot: Decoupling Pixel Detail from Global Attention","SpecSlot treats tokenization as allocation: only each anchor's lowest-frequency slot joins global attention. The paper reports 9.14 FID on ImageNet.","Pixel-space generative models learn directly from noise to pixels, sidestepping the separately trained autoencoder. The price is that a full-resolution image has to be represented as Transformer tokens. Most such models spend one token per spatial patch, which means a single grid decides three things at once: what the model reads, what each token predicts, and how many tokens join global attention. Want finer detail and the attention bill goes up with it. On a 256×256 image, dropping the patch size from 16×16 to 8×8 takes the token count from 256 to 1,024 and makes the quadratic term of dense attention 16× larger.\n\nThe problem is that coupling. What a token is allowed to communicate globally is decided by where it sits in the grid, not by what it finally produces.\n\n## Rewriting tokenization as budget allocation\n\nA team from Peking University, Nanyang Technological University and SenseTime proposes SpecSlot, and it changes the split itself. The total image token budget is written as N = W × K: W is spatial anchor density, K is spectral subdivision inside each anchor. Every anchor expands into K slots, and each slot owns a fixed, disjoint band of DCT coefficients. A spatial residual is then added on the output side as a local correction. The generative state, the target and the loss all stay in pixel space.\n\nCommunication is assigned by role. In each anchor, the lowest-frequency slot is the carrier, and only that slot enters global attention; every other slot interacts locally inside its own anchor. The schedule is global, global, local, local, repeated three times, giving six global blocks and six local blocks of identical dimensions. The global sequence length therefore tracks W instead of W×K. In the ImageNet-256 configuration with s=16 and K=4, the global sequence holds W+C = 288 tokens, where a scheme that puts every slot into global attention needs 1,056. The corresponding quadratic term drops from 12(WK+C)²d to 6(W+C)²d + 6WK²d.\n\nThe paper summarizes this as tokenization as allocation. It loosens two things that were previously welded together: input density is set by the input grid, and output role is set by the frequency band inside an anchor. Neither has to hold the other hostage.\n\n## The advantage only opens up at 512 resolution\n\nIn the authors' own evaluation, SpecSlot W256\u002FK4 scores 6.59 FID at 129.20 GFLOPs on ImageNet-256, only 0.09 better than the compute-matched JiT-CM control at 6.68. The paper says that gap sits inside the 95% interval across five random seeds, so it is essentially a tie. The real quality leader is still the more expensive JiT-B\u002F8 at 4.34 FID and 218.50 GFLOPs.\n\nAt ImageNet-512 the gap opens. W256\u002FK4 reaches 9.14 FID at 135.64 GFLOPs, 2.84 FID below the compute-matched control at 11.98 while using 3.9% fewer forward FLOPs. Against the costlier JiT-B\u002F16 at 8.45 FID and 219.56 GFLOPs, the gap is only 0.69 FID, at 38.2% less compute and 12.69 against 7.93 img\u002Fs throughput. The authors' explanation is arithmetic: at 512 each anchor covers four times as many pixels as at 256, while the global sequence stays fixed at 288 tokens, so the extra detail stays inside the anchor.\n\nFor text-to-image at 256 resolution, DPG-Bench comes in at 62.21, 1.06 points above JiT-B\u002F8 at 61.15, with forward compute dropping from 218.77 to 129.86 GFLOPs, a 40.6% reduction. The same model scores 26.99 on GenEval, below JiT-B\u002F8 at 29.78. The architecture transfers to text conditioning without structural changes, and the price is giving up ground on a different alignment metric.\n\n## A detail worth not skipping: matching compute is not matching speed\n\nThe routing comparison table says the most about what this work actually is. In the same W256\u002FK4 configuration, GGLL gives 6.59 FID at 129.20 GFLOPs and 9.06 img\u002Fs. The compute-matched dense-global control Dense-global-CM gives 6.81 FID at 128.96 GFLOPs but only 5.80 img\u002Fs. Measured forward GFLOPs are nearly identical, yet wall-clock throughput is far apart. Across sampling budgets of 20, 30 and 50 Heun intervals, GGLL throughput runs 1.47 to 1.49× the control.\n\nOne level further up, Full-global puts all 1,024 slots into global attention and posts the table's lowest FID at 5.84, at a cost of 225.82 GFLOPs, 74.8% above GGLL. So SpecSlot does not take the top score at equal compute. It is cutting 42.8% off the cost of routing everything globally while still holding a better quality than the compute-matched control. It picks the middle operating point, the one constrained by attention, not the leftmost one.\n\n## Mechanistic evidence: the low frequencies really do come back first\n\nThe paper supplies mechanistic evidence. Holding W256\u002FK4\u002FB1536 fixed and changing only which slot gets cross-anchor access: slot 0 gives 6.59, and swapping in slot 1, 2 or 3 gives 6.91, 7.21 and 6.97, all worse. Output-branch probes show the spectral term carries the larger self-energy, while the spatial term's share grows from 0.07 to 0.23 across the flow. Removing either branch at t=0.90 raises velocity MSE by 32.1× and 64.1× respectively, and at t=0.05 by 1.1× and 2.1×. The spectral scaffold is there from the start; the spatial correction is high-frequency top-up that arrives late.\n\nAllocation ablations show that at a fixed slot budget of 2,048, more anchors means a lower score (7.28 for W64\u002FK32, 6.28 for W256\u002FK8, 5.66 for W1024\u002FK2). With anchors fixed at 256, raising K from 2 to 4 to 8 gives shrinking returns, and the last doubling buys only 0.31 FID.\n\n## My read\n\nAll of these numbers currently come from a single origin: the paper, the official repository and the project page belong to the same team, every score is self-reported, and no third party has reproduced it. The paper itself is listed as under review at ICLR 2027. The right way to read it is as a methodological proposal, not as an established result.\n\nThe direction is still worth watching. Most of the past two years of pixel-space generation discussion has circled around how to remove the autoencoder entirely. This work swaps the question for how to decouple detail from communication budget. It does not beat the costlier JiT-B\u002F8 at 256 resolution; what it wins is a non-dominated operating point in the constrained regime. The advantage clearly appears at 512, which hints that the payoff of decoupling grows together with resolution instead of spreading evenly.\n\nFor people building generative systems, the more immediately useful takeaway may be the side observation about controls: two models matched on compute can differ substantially in wall-clock speed. Cutting FLOPs is not the same as cutting latency, especially when what you actually change is the shape of attention rather than only its scale. What to watch next is whether this band division holds at larger scale, and whether the split between low and high frequency still makes sense on temporal data like video.\n\nPaper and code: abstract page [arXiv:2610.40647](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40647), project page [yangjiongfan.site\u002Fspecslot](https:\u002F\u002Fyangjiongfan.site\u002Fspecslot\u002F), repository [JiongfanYang\u002FSpecSlot](https:\u002F\u002Fgithub.com\u002FJiongfanYang\u002FSpecSlot).\n","specslot-pixel-slot-global-attention-decoupling","2026-10-31T05:20:00Z","2026-10-31T05:16:22.390533Z","2026-10-31T05:16:22.390540Z",true,"agent",87,[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},"b270aaf5-e8a5-40a2-98c7-612acc3573ed","任意顺序图像生成不用重训:适配1.25倍算力换来补图扩图","any-order-image-generation-ar-adaptation","2026-10-31T07:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b369ce1a-f8a3-4aba-ba78-c73c097a413d","MemT-1.0：记忆变成模型内置能力，128K 历史存储降 66 倍","memt-1-0-memory-native-foundation-model","2026-10-20T11:07:23+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"2266cea6-06f1-4932-8905-1bc3f2e5a8c0","Meta FAIR 字节蒸馏研究:End-Of-Token 渐近反超 token 蒸馏 4%,数据只需 1\u002F6","meta-fair-byte-distillation-token-ceiling-2026-09","2026-09-15T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"3d922c00-afcb-4f1c-a6d5-8f9d6c10c642","从 Kimi Linear 到 Kimi K3:MoE 推理效率战里被忽略的架构升级","kimi-k3-latentmoe-kda-attnres-nope","2026-07-30T00:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b5638cab-a4d6-44ac-9230-32ed0a4cba9d","ARMT 把「记忆」焊进 Transformer:用恒定显存换无限上下文","armt-associative-recurrent-memory-transformer","2026-07-23T00:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d2844cbb-b70e-469c-95b8-71cee8d735a6","给 Transformer 装上「CNN 鼻子」:用 0.01% 的参数量换 benchmark 普涨","transformer-cnn-nose-0-01-percent","2026-07-22T12:10:00+00:00"]