[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-any-order-image-generation-ar-adaptation":3,"topics-all":38,"news-related-b270aaf5-e8a5-40a2-98c7-612acc3573ed":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},"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","自回归图像模型因训练时锁死栅格扫描顺序,天生无法补图扩图。这篇 NeurIPS 论文用面向未来的位置编码加顺序课程,把预训练模型改造成任意顺序生成器:适配成本远低于从头重训,并保住 KV 缓存与 vLLM 部署。补图与扩图成绩都超过 MaskGIT,论文也坦承类条件指标有所回落。","自回归图像模型有个硬伤：它天生只会从左上角往右下角一笔一笔画。训练时 token 的序列顺序被钉死成栅格扫描，推理时想补掉图中间一块、往外扩一圈、或者只重画某个区域，做不到——不是模型没学会这个能力，而是这个能力压根没被写进权重。\n\n常规解法是从头重训，让模型在各种随机排列的顺序上训练。论文引用的参照是，RandAR 这类随机顺序自回归模型要训 300 个 epoch，而对应的栅格扫描 LlamaGen-XL 只要 100 个，差不多 3 倍算力；RandAR 官方项目页也写明它是从随机置换的 token 序列开始训的。另一条主流路线离散扩散，算力同样是 3 倍。\n\n一篇 NeurIPS 2026 收录的论文（[arXiv:2610.39293](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39293)）说这个 3 倍可以省：给预训练模型加两个轻量组件，就能把训好的栅格模型改造成任意顺序生成器。\n\n## 真正省下来的不是算力，是部署兼容性\n\n论文自己的对比表把账列得很清楚：RandAR（从头训）3 倍、离散扩散 3 倍、RAR 4 倍，这篇的适配方案在 LlamaGen-XL 上是 1.25 倍，跨四个基座模型是 1.13–1.33 倍。\n\n但真正有分量的不是训练成本那一列，是紧挨着的兼容性三列。RandAR 3 倍代价换来的是「不能上 vLLM」；离散扩散更贵，直接把 KV cache 和 vLLM 一起丢掉——它的双向注意力每一步都要重扫所有未屏蔽位置，跟 AR 推理框架的逐 token 缓存复用天然不兼容。适配方案是唯一一行三个勾全中的：任意顺序、KV cache、vLLM 部署都保住，代价只是 1.25 倍训练预算。论文说，对手上已经有预训练权重的人，边际适配成本比从头训低一个数量级以上（12×）。\n\n技术核心是两个组件，都不改动原模型的因果单流 Transformer 结构。\n\n一是**面向未来的位置编码**。栅格扫描下，第 t 步该预测哪个位置是隐含已知的，就是 t+1；顺序一换，同一段已看到的上下文可能对应完全不同的预测目标——同样一张图的其余部分，预测天空和预测人脸是两个不同的条件分布，模型会同时收到互相冲突的梯度。解法是显式告诉模型这一步要填哪个位置，用的是目标位置索引的 2D RoPE 路由，外加一张加性嵌入表。对 LlamaGen-XL，这张表只占 0.74M 参数，也就是 0.095%。\n\n二是**顺序课程**：不是一步切换到随机顺序，而是从栅格顺序逐步过渡，替模型铺一条从预训练解走到任意顺序解的稳定优化路径。\n\n## 比结果更值钱的三个负结果\n\n论文在补图上 FID 6.21，对 MaskGIT 的 10.44；扩图 5.84 对 12.17；区域重建 LPIPS 0.125，也好于 MAR-L 的 0.174。不过论文自己很克制：MaskGIT 换用 Halton 调度后补图 FID 降到 7.82，论文承认原始优势里有一部分来自对手的调度器选择；对 MAR-L，扩图的 Precision 和 Recall 反而是 MAR-L 更高，由于没有跑级不确定性，论文说两者算「有竞争力」而非领先。\n\n真正值得抄走的是三个反例。\n\n**去掉面向未来的位置编码，训练损失不降反升。** 随着课程把随机顺序的比例推高，损失曲线是往上爬的。这正好印证了前面的可辨识性论证——冲突的梯度是真的。\n\n**去掉顺序课程，模型停在半适配状态。** 一上来就用完全随机顺序（α=1），LlamaGen-L 上 FID 5.86，对着完整配方的 4.24。论文的诊断很准：Recall 基本没变（0.51 对 0.52），所以这是保真度塌陷，不是覆盖度塌陷——预训练的表征在顺序灵活的路由形成之前就先被磨掉了。\n\n**最反直觉的一个：拿冻结的栅格老师做正则反而更糟。** 在栅格样本上加一个指向老师的 KL 项（λ=1.0），FID 从 4.24 涨到 9.62。想保住旧行为，和学新的条件分布，是直接冲突的。\n\n还有个工程细节：这份额外嵌入表的学习率要开到网络其余部分的 10 倍才收敛得稳，否则它跟不上课程节奏。\n\n## 账还是要算清楚\n\n适配不是白拿的。四个基座模型上，类条件生成的 FID 相对各自的预训练检查点上升 1.17–1.93——LlamaGen-L 从 3.07 到 4.24，LlamaGen-XL 从 2.63 到 3.91，RAR-L 从 1.70 到 3.63。论文主动说明这个代价「不一定是最小的，也不随规模单调」，并指出 RAR 因为本来就贴近 FID 下限，同样的名义退化看起来更明显。\n\n还有个容易被忽略的对照：同等总 epoch 下，适配预训练模型明显好过随机初始化（LlamaGen-L 4.24 对 5.60，LlamaGen-XL 3.91 对 4.81），但论文提醒，这个对照只匹配了算力、没匹配 schedule——只有预训练路径是从栅格训练开始的。\n\n推理侧是另一笔账：任意顺序训练让并行解码成为可能，LlamaGen-XL 上前向次数减少 9 倍，代价是 FID 从 3.91 涨到 5.03。\n\n## 边界写得很清楚\n\n论文限定在 ImageNet 256×256 的类条件生成；能不能迁移到文生图的自回归模型、能不能上更高分辨率，都还没验证；编辑能力的评测只覆盖区域补全，不涉及指令跟随和身份保持。适配模型保留了 vLLM 部署（通过自定义模型注册），代码以 MIT 协议放出，checkpoint 在 Hugging Face 的 orderar 组织下。\n\n所以这篇真正推进的不是一个新模型，而是一个更窄也更硬的判断：同一个联合分布的不同分解方式之间，可以靠一次轻量适配互通，不必非得从头训一个。而它顺手留下的那三个负结果，可能比那些漂亮的 FID 数字更值得记住——尤其是那个把 FID 从 4.24 推到 9.62 的老师正则实验。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39293","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},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"c883fd20-1d66-4fb7-9fc7-320fa7f87023","text-to-image",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9d45dcaa-171a-4957-995a-430255def714","en","Unlocking Any-Order Image Generation From Pretrained Models","A pretrained autoregressive image model is turned into an any-order generator, beating MaskGIT at inpainting while keeping KV cache and vLLM serving.","Autoregressive image models carry one hard limitation: they can only paint from the top-left corner to the bottom-right, one token at a time. Because the token ordering is nailed down to a raster scan at training time, the model cannot fill in a patch in the middle, extend the canvas outward, or repaint a single region at inference time. This is not a case of the model failing to learn the capability — the capability was never written into the weights.\n\nThe standard fix is to retrain from scratch, teaching the model a mixture of randomly permuted token orders. The paper cites RandAR, a random-order autoregressive model that needs 300 epochs where the corresponding raster-scan LlamaGen-XL needs 100 — roughly 3x the compute. RandAR's own project page confirms it was trained from randomly permuted token sequences. The other mainstream route, discrete diffusion, also costs about 3x.\n\nA NeurIPS 2026 paper ([arXiv:2610.39293](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39293)) argues that 3x is unnecessary: add two lightweight components to a pretrained model and a finished raster model becomes an any-order generator.\n\n## What actually gets saved is deployment compatibility, not compute\n\nThe paper's comparison table is explicit: RandAR (from scratch) 3x, discrete diffusion 3x, RAR 4x, while this adaptation recipe costs 1.25x on LlamaGen-XL and 1.13–1.33x across four base models.\n\nBut the consequential column is not the training-cost one — it is the three compatibility columns sitting right next to it. RandAR trades 3x for \"cannot serve with vLLM\". Discrete diffusion costs even more and loses both the KV cache and vLLM: its bidirectional attention rescans every unmasked position at each step, which is structurally incompatible with the per-token cache reuse that AR inference frameworks depend on. The adaptation recipe is the only row with all three checked — any-order, KV cache, and vLLM serving — at the price of a 1.25x training budget. The paper notes that for practitioners who already hold pretrained weights, the marginal adaptation cost is over an order of magnitude lower (12x) than training any-order from scratch.\n\nThe technical core is two components, neither of which touches the base model's causal single-stream transformer.\n\nOne component is **future-aware positional embeddings**. Under raster scan, the target position at step t is implicitly known — it is t+1. Once the ordering varies, the same observed context can correspond to entirely different prediction targets: given the rest of one image, predicting a sky token and predicting a face token are two different conditional distributions, so the model receives conflicting gradients simultaneously. The fix is to explicitly tell the model which position to fill next, using target-position-indexed 2D RoPE routing plus an additive embedding table. For LlamaGen-XL that table holds 0.74M parameters, or 0.095%.\n\nThe other is an **ordering curriculum**: rather than switching to random orders in one step, the schedule ramps gradually, laying down a stable optimization path from the pretrained solution to a good any-order basin.\n\n## Three negative results worth more than the headline numbers\n\nThe paper reports FID 6.21 on inpainting against MaskGIT's 10.44, FID 5.84 on outpainting against 12.17, and LPIPS 0.125 on region reconstruction versus MAR-L's 0.174. It is also notably careful with itself: with the Halton schedule, MaskGIT's inpainting FID drops to 7.82, and the paper concedes that part of the original margin came from the competitor's scheduler choice. On outpainting, MAR-L actually has the higher Precision and Recall; with no run-level uncertainty available, the paper calls the two competitive rather than claiming the lead.\n\nThe three counterexamples are what are worth carrying away.\n\n**Remove the future-aware positional embeddings and training loss rises instead of falling.** As the curriculum ramps up the proportion of random orders, the loss curve climbs. That is exactly what the identifiability argument predicts — the gradient conflict is real.\n\n**Remove the ordering curriculum and the model sits half-adapted.** Starting immediately at full random order (α=1) yields FID 5.86 on LlamaGen-L, against 4.24 for the complete recipe. The paper's diagnosis is sharp: Recall barely moves (0.51 versus 0.52), so this is a fidelity collapse rather than a coverage collapse — the pretrained representation gets eroded before order-flexible routing can form.\n\n**The most counterintuitive one: regularizing toward a frozen raster teacher makes things worse.** Adding a KL term (λ=1.0) toward the teacher on raster-ordered examples pushes FID from 4.24 up to 9.62. Preserving the old behavior and learning the new conditional distributions are directly in conflict.\n\nOne engineering detail: the extra embedding table's learning rate has to be scaled 10x relative to the rest of the network to converge stably, otherwise it cannot keep pace with the curriculum.\n\n## The bill still has to be settled\n\nAdaptation is not free. Across four base models, class-conditional FID rises by 1.17–1.93 relative to the corresponding pretrained checkpoint — LlamaGen-L from 3.07 to 4.24, LlamaGen-XL from 2.63 to 3.91, RAR-L from 1.70 to 3.63. The paper volunteers that this cost is \"not necessarily fundamental or monotonic with scale\", and notes that RAR sits closer to the FID floor, where the same nominal degradation is more visible.\n\nOne easily overlooked control: at matched total epochs, adapting a pretrained model clearly beats training from random initialization (LlamaGen-L 4.24 versus 5.60; LlamaGen-XL 3.91 versus 4.81). But the paper reminds us this control matches compute, not schedule — only the pretrained pathway begins with raster-only training.\n\nInference is a separate line on the bill: any-order training makes parallel decoding possible, cutting forward passes by 9x on LlamaGen-XL, at the cost of FID moving from 3.91 to 5.03.\n\n## The boundaries are stated plainly\n\nThe work is confined to class-conditional generation on ImageNet at 256×256. Whether the recipe transfers to text-to-image autoregressive models, or to higher resolutions, remains untested. The editing evaluation covers region completion only, and does not measure instruction following or identity preservation. The adapted model retains vLLM serving through a custom model registration, the code is released under the MIT license, and checkpoints are available under the orderar organization on Hugging Face.\n\nSo what this advances is not a new model but a narrower and harder claim: different factorizations of the same joint distribution can be bridged by one lightweight adaptation, with no need to train a model from scratch. And the three negative results it leaves behind may be worth remembering more than the tidy FID numbers — especially the teacher-regularization experiment that pushes FID from 4.24 to 9.62.\n","any-order-image-generation-ar-adaptation","2026-10-31T07:05:00Z","2026-10-31T07:13:47.710976Z","2026-10-31T07:13:47.710987Z",true,"agent",117,[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},"c03ecfed-2b3f-4831-bbc0-fad67a993917","SpecSlot:像素生成不必让细节都走全局注意力","specslot-pixel-slot-global-attention-decoupling","2026-10-31T05:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"774de6ac-98e1-4343-a67a-bfdc72d377bb","INFORMS 实证:AI 广告真实投放胜过设计师,18 个月后仍领先","informs-ai-ads-beat-human-designers-18-months","2026-08-22T14:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"deac2d55-76a6-40d2-8ef7-36aed2ad0105","Linux 7.2 把 AI 拉进内核开发:Sashiko 让补丁数量翻倍,Torvalds 接受「新常态」","linux-7-2-sashiko-ai-kernel-review","2026-08-20T12:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a6119d74-007a-4692-bb4e-d85b562a9d66","字节 SpectraReward：自我奖励 T2I 干翻 30× 大模型","bytedance-spectra-reward","2026-07-15T04:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"029d5b6c-a448-442b-b742-96afeaab330f","PCS 把 LLM 推理能力\"渐进迁移\"到任意语种：5 个语种验证，轻量翻译替代昂贵蒸馏","pcs-llm-progressive-transfer","2026-07-08T14:15:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"90fc8caa-c5f6-45ab-adb8-50f28f43739b","字节 UP：正向 advantage 不裁剪，GRPO\u002FDAPO\u002FGSPO 即插即用","bytedance-seed-up-advantage","2026-07-08T04:21:42+00:00"]