[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-refdecoder-reference-conditioned-video-decoder-zh":3,"article-related-refdecoder-reference-conditioned-video-decoder-zh":36,"series-research-001e062e-f246-4bf0-aa04-27506febcf7b":88},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":18,"translated_content":10,"views":19,"is_premium":20,"created_at":21,"updated_at":21,"cover_image":11,"published_at":22,"rewrite_status":23,"rewrite_error":10,"rewritten_from_id":24,"slug":25,"category":26,"related_article_id":27,"status":28,"google_indexed_at":29,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":30,"topic_cluster_id":34,"embedding":35,"is_canonical_seed":20},"001e062e-f246-4bf0-aa04-27506febcf7b","RefDecoder 讓影片解碼器吃參考圖","\u003Cp data-speakable=\"summary\">RefD\u003Ca href=\"\u002Fnews\u002Flovable-backs-atech-vibe-coding-hardware-zh\">ecod\u003C\u002Fa>er 讓影片解碼器直接參考輸入影像，改善重建細節與畫面一致性。\u003C\u002Fp>\u003Cp>在影片生成和編輯流程裡，條件訊號通常集中在去噪網路，真正把 latent 還原成畫面的 decoder 卻常常是無條件的。這篇論文認為，這種設計落差，正是生成影片容易丟細節、也容易在重建時偏離原始影像的原因之一。\u003C\u002Fp>\u003Cp>作者提出的解法叫 RefDecoder。它不是去重寫整個影片生成系統，而是把參考影像直接注入解碼路徑，讓 decoder 在還原畫面時也能看到高保真 reference frame。對做 image-to-video、影片編輯，或任何需要貼近原圖的工作流來說，這種做法很實際：不用整套重訓，也有機會把輸出品質往上拉。\u003C\u002Fp>\u003Ch2>這篇在修哪個痛點\u003C\u002Fh2>\u003Cp>這篇論文盯上的，是目前影片生成堆疊裡一個很典型的結構問題。latent diffusion 類方法通常會把大部分條件控制放在 denoising network，讓模型在去噪時盡量遵守輸入影像或提示詞；但到了 decoder 這一步，條件訊號卻常被拿掉。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912646805-czy9.png\" alt=\"RefDecoder 讓影片解碼器吃參考圖\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>問題就在這裡。decoder 才是把細節真正補回來的地方。如果這一段不看 reference，它就只能靠 latent 自己猜，結果常見的就是結構變形、紋理糊掉，或是和原始影像的內容慢慢走樣。論文把這種不對稱視為 detail loss 和 inconsistency 的來源。\u003C\u002Fp>\u003Cp>RefDecoder 的目標，就是把這個缺口補起來。它不把 decoding 當成單純的還原步驟，而是把參考影像當成 decoder 的一部分輸入，讓還原過程一路都能對照原圖。\u003C\u002Fp>\u003Ch2>RefDecoder 怎麼做\u003C\u002Fh2>\u003Cp>它的核心概念其實很直白：在 denoised video latents 之外，再把 reference image 一起送進 decoder。論文使用的是 reference attention，讓 decoder 在每個 up-sampling stage 都能同時處理兩種訊號。\u003C\u002Fp>\u003Cp>更具體一點，系統先用一個輕量的 image encoder，把參考影像轉成高維 \u003Ca href=\"\u002Fnews\u002Fatlas-one-token-visual-reasoning-zh\">toke\u003C\u002Fa>ns。接著，這些 tokens 會和影片 latent tokens 在 decoder 裡合併。這樣一來，decoder 在補細節時不是只靠 latent 猜測，而是有原始 reference 可以對照，能更穩定地恢復結構與紋理。\u003C\u002Fp>\u003Cp>這個設計的重點，是它把改動侷限在 decoder 端。論文明講，RefDecoder 可以直接插進既有的影片生成系統，而且不需要額外 fine-tuning。這代表它比較像一個可替換元件，而不是要你把整條訓練管線砍掉重來。\u003C\u002Fp>\u003Cp>對工程團隊來說，這種局部升級很有吸引力。你可以保留原本的 latent video generator，只調整最後的解碼階段，先看畫質和一致性有沒有改善，再決定要不要進一步改架構。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>根據摘要，作者把 RefDecoder 放到多個 decoder backbone 上測試，包括 Wan 2.1 和 VideoVAE+，結果都看到一致改善。這表示方法不是只對單一架構有效，而是有一定的可移植性。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912633315-buni.png\" alt=\"RefDecoder 讓影片解碼器吃參考圖\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>論文也提到，它在多個 reconstruction benchmarks 上表現更好，包括 Inter4K、WebVid 和 Large Motion。不過摘要沒有公開完整 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 表格，所以目前能確認的只有這些資料集名稱與整體趨勢，細部數字沒有完整展開。\u003C\u002Fp>\u003Cp>摘要裡唯一明確的量化數字，是相較於 unconditional baselines，PSNR 最多提升 +2.1 dB。PSNR 是重建品質常看的指標，這個結果代表 RefDecoder 在像素層級上更能貼近 reference input，也就是說畫面更不容易在還原時失真。\u003C\u002Fp>\u003Cp>除了重建，作者也說它在 V\u003Ca href=\"\u002Fnews\u002Fentitybench-long-range-video-consistency-zh\">Benc\u003C\u002Fa>h I2V 上有更好的成績，尤其是在 subject consistency、background consistency 和 overall quality 這幾項。這點很重要，因為 image-to-video 不只是要單張帧清楚，還要讓主體和背景在時間上維持穩定。\u003C\u002Fp>\u003Cp>另外，論文還提到 RefDecoder 能 generalize 到 style transfer 和 video editing refinement。這表示它不只是一個單一任務的小技巧，而是可能對多種需要保留來源內容的影像生成任務都有幫助。不過摘要沒有提供這些延伸場景的獨立數字，所以目前只能把它視為能力上的延伸，而不是已被完整量化的結論。\u003C\u002Fp>\u003Cul>\u003Cli>可套用到 Wan 2.1、VideoVAE+ 等不同 decoder backbone\u003C\u002Fli>\u003Cli>在 Inter4K、WebVid、Large Motion 上都有改善\u003C\u002Fli>\u003Cli>PSNR 最多提升 +2.1 dB\u003C\u002Fli>\u003Cli>VBench I2V 的 subject consistency、background consistency、overall quality 也更好\u003C\u002Fli>\u003Cli>宣稱可直接替換進既有系統，不需要額外 fine-tuning\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>對開發者有什麼影響\u003C\u002Fh2>\u003Cp>如果你在做影片生成、image-to-video，或是影片編輯工具，這篇論文的訊號很明確：decoder 不是只是把 latent 轉成像素的最後一步，它本身就是影響品質的關鍵位置。很多團隊會把注意力放在 denoiser 或 prompt conditioning，但這篇提醒你，最後的重建階段也一樣重要。\u003C\u002Fp>\u003Cp>這對需要高度貼近來源影像的工作流特別有用。像是產品展示、角色動畫、style transfer、影片修飾這類場景，只要結構或背景有一點偏差，使用者就會很快看出來。RefDecoder 的思路，是把 reference 直接帶進最後一段，降低這種偏移。\u003C\u002Fp>\u003Cp>另一個實務優點，是它主打不用額外 fine-tuning。對工程團隊來說，這意味著可以先把它當成 decoder-level 的替換方案來評估，而不是先投入大規模重訓成本。若現有系統已經穩定，這種 drop-in 式的改動通常更容易進行 A\u002FB 測試。\u003C\u002Fp>\u003Cp>但這篇摘要也留下不少工程面問題。它沒有提到 runtime cost、記憶體開銷、延遲影響，也沒有說 reference attention 會不會讓推論更重。對實際部署來說，這些資訊很關鍵，因為畫質提升如果換來太高成本，未必適合線上服務。\u003C\u002Fp>\u003Cp>摘要也沒有把所有 benchmark 的完整數字公開，所以目前只能看出方向是正面的，還不能直接推論它在每個資料分佈上都同樣有效。實作上，團隊還是得用自己的內容分佈去驗證，尤其是主體類型、動作幅度、背景複雜度不同時，效果可能會有差。\u003C\u002Fp>\u003Ch2>這篇論文真正的意義\u003C\u002Fh2>\u003Cp>RefDecoder 的價值，不在於它提出了一個很複雜的新生成框架，而在於它指出一個常被忽略的事實：條件控制如果只做一半，最後的 decoder 還是會把資訊弄丟。把 reference conditioning 往後推到解碼端，可能就是補上畫面一致性的關鍵一步。\u003C\u002Fp>\u003Cp>對\u003Ca href=\"\u002Ftag\u002F台灣開發者\">台灣開發者\u003C\u002Fa>來說，這類研究很值得注意，因為它比較接近「怎麼把現有系統做得更穩」而不是「重新發明一套模型」。如果你已經在用 latent video generation 流程，這種局部模組升級的思路，往往比全面重建更容易落地。\u003C\u002Fp>\u003Cp>總結來看，RefDecoder 想解決的是影片生成裡的最後一哩路問題：不是讓模型更會猜，而是讓它在把畫面還原出來的那一刻，還記得原圖長什麼樣。對重建、編輯和 image-to-video 來說，這個差異可能比想像中更大。\u003C\u002Fp>","RefDecoder 把參考圖直接送進影片解碼器，補上傳統流程只在去噪端做條件控制的缺口，目標是提升重建細節、一致性與可用性。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15196",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912646805-czy9.png",[13,14,15,16,17],"video generation","video decoder","reference conditioning","image-to-video","PSNR","zh",2,false,"2026-05-16T06:23:33.170076+00:00","2026-05-16T06:23:32.963+00:00","done","50da1a2f-5cd6-48d9-8aa5-e75b6add633b","refdecoder-reference-conditioned-video-decoder-zh","research","66608799-65b1-4143-afc1-d1457cdd696a","published","2026-05-16T09:00:16.708+00:00",[31,32,33],"RefDecoder 把參考影像直接送進影片解碼器，補上傳統流程只在去噪端做條件控制的缺口。","摘要中最明確的數字是 PSNR 最多提升 +2.1 dB，並提到 VBench I2V 的一致性與品質更好。","方法主打可插入既有系統、且不需要額外 fine-tuning，但摘要沒有提供 runtime、記憶體或完整 benchmark 表格。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.03292327,-0.006174568,0.012791867,-0.11217382,-0.014532306,0.01834283,-0.018333822,-0.02418654,0.030412707,-0.021155417,0.011212082,-0.052688807,0.010443596,0.0266896,0.1163732,0.0043267994,0.0060468167,0.002974068,-0.0034499734,-0.0029313574,0.00062022096,0.018456206,0.0074437438,0.015082636,-0.0047606495,0.008382282,0.02155934,0.017226335,0.054772798,-0.0013017718,0.033427816,0.0009553193,0.024680087,0.027134826,0.0026876654,0.013552406,0.00077384606,0.00610979,0.013661993,0.01846722,-0.030599095,-0.026323264,0.019914357,-0.003921484,-0.010548297,0.009680533,0.010334064,-0.03326468,0.0050598546,0.0065294285,0.016550917,0.02581431,-0.03140669,-0.15999267,0.0038526591,0.03252856,0.009951926,-0.0056309416,-0.011349095,0.008409699,-0.00086511706,-0.0008699836,-0.003320081,-0.037098117,-0.011085604,-0.014854504,0.0036904763,0.0032296418,-0.009727746,-0.0025205263,-0.010901088,0.023633076,-0.013762413,0.023697674,0.044369478,-0.011705963,-0.01328177,0.015890058,0.014341451,0.03597275,0.0007906677,-0.002463703,0.0075560776,0.0014194178,0.02272967,0.020911463,-0.019112343,0.008837329,0.0004850175,0.00085139833,0.0006945895,0.006482859,0.032506645,0.014193995,0.009293257,0.017528716,-0.015674772,-0.0134087475,0.006821831,0.0032093574,-0.0025565205,-0.022333005,0.025612028,0.016416514,0.02751147,-0.008715392,-0.008042947,0.01322827,-0.00961655,0.016056042,-0.01601625,-0.024187377,-0.015870294,-0.0008831635,-0.007943682,-0.13035029,-0.004474905,0.04066318,-0.0063036187,0.02427862,0.008967158,-0.0026186814,-0.010697308,0.009411636,0.013289513,-0.01179971,0.020951234,0.0018933788,-0.032983907,-0.0035221127,-0.0024507043,-0.00085346994,0.0009944815,-0.017393284,0.0028644996,0.008801692,0.015520513,-0.023451198,-0.016904294,0.0026126348,0.002838951,-0.0038269518,-0.011292523,-0.017581211,-0.028178325,-0.028343191,-0.04370299,0.00018470373,-0.0068260827,-0.028728504,0.013618552,-0.0093128355,0.012323083,-0.0038772102,0.039915714,-0.024370013,-0.009850517,0.00298715,0.0007134741,0.011113891,0.0057931487,-0.005263469,0.02370072,-0.009163378,0.007845087,0.018513344,0.015530259,0.0014763968,-0.002737041,0.01661336,-0.016259152,0.0063442937,-0.0068578203,-0.015321076,0.01308169,0.020806786,-0.007093801,-0.026480807,-0.009475694,-0.01813454,0.0021473977,-0.0017236042,-0.0047858874,0.004356693,0.03551901,-0.008320309,0.0025888009,-0.008619595,0.030511219,0.003698042,-0.0018544592,-0.010088973,0.016578475,-0.007537329,0.004282783,-0.021101886,0.0014081196,0.008624136,0.031268943,0.02809636,-0.022347648,-0.00028244857,0.015217534,-0.006105496,-0.007255999,-0.002847646,0.012396448,0.004777489,0.008933226,0.0071158432,0.006182148,0.0026815399,0.009757286,-0.012120678,-0.006733808,-0.0052863862,-0.010153017,-0.02486055,-0.024243576,-0.033026844,0.021659624,-0.016400289,-0.001978493,0.00320024,-0.016499912,-0.05428664,0.0028648705,0.004379867,0.016997483,-0.0048053325,-0.009739834,0.0004449209,-0.0009278734,-0.0142527595,-0.0063756495,0.0021541477,-0.009159944,0.03261712,0.05135424,0.00798285,-0.027450403,0.010744465,0.008854655,-0.0003682504,0.015973093,0.019543327,-0.0014628599,0.004812626,0.016311796,0.013208945,-0.008922144,-0.0057951827,-0.016022755,-0.030946547,-0.011701387,0.009441017,0.001985446,-0.00019069694,-0.0068548927,-0.0031475883,-0.010752752,0.013160078,-0.0061731697,0.018734941,0.05593785,0.018354151,0.010943232,0.008635164,-0.025248354,0.011135715,0.03199882,0.012069732,0.018564124,-0.020286882,0.017196046,-0.013714602,-0.03546325,0.047863454,0.009178836,-0.040833317,-0.016027842,0.023670407,0.0070986627,0.019218795,0.016023286,0.013185714,0.006133391,-0.0561061,0.004794383,-0.018257832,-0.011317934,0.022029381,0.007188654,-0.017726053,0.011463414,-0.036476232,-0.0015942777,-0.006059042,0.016063966,0.020596193,0.012572148,0.02558933,0.003093832,0.018486368,0.014940606,-0.0011129993,-0.020088533,0.038742963,0.016683768,0.010216734,-0.017245483,0.0029046484,-0.010699573,-0.0047698603,-0.021190293,-0.0038079468,0.015730863,0.0067567094,-0.001091139,-0.00949894,-0.01168959,-0.007678187,-0.030930351,0.009961128,-0.013823266,-0.0005205573,-0.024699213,-0.032483187,-0.0073637124,-0.0067099445,-0.02561862,0.022504684,0.01145023,-0.017578203,0.008069318,-0.0023678187,0.021822887,-0.002867794,-0.039981388,-0.008444229,0.019421633,0.00431082,-0.008874898,0.0062848884,-0.026749833,0.018830707,0.005342438,0.0016990384,0.013227943,-0.04342751,-0.0030651754,0.0064109014,0.0012226121,-0.013416639,-0.0037314862,-0.002317421,-0.01928018,-0.02951134,0.020217698,-0.009049079,0.00833141,0.010200491,-0.015246018,-0.011429364,-0.003477018,-0.00081796217,-0.0029825186,0.0020043228,-0.013036861,-0.0076357005,0.0145985,-0.008080732,-0.00036643283,-0.0010498244,-0.02369714,0.0028454552,-0.0064688036,-0.0082621705,-0.0151557615,0.006013825,0.014715293,0.012676981,-0.024693822,-0.029809104,-0.017274337,-0.0015892423,0.022399213,-0.02675442,-0.016797049,0.01240908,0.0032658,-0.018028256,0.0009247349,0.018598963,0.024096197,0.0064328643,-0.0019257093,0.010355492,-0.00054279383,-0.0034730004,-0.000541823,0.011889351,-0.0047208527,-0.010877129,0.015276449,0.016487569,-0.0033002647,0.013175833,-0.020287868,-0.008792949,0.016027905,-0.030739369,0.034970243,-0.014064054,0.016984034,0.00931327,-0.015742023,-0.011980087,0.014018685,-0.0056538773,0.0040415656,-0.035278026,0.0014098239,0.024714125,-0.0106815165,0.012908874,-0.013676399,-0.007036915,0.033354826,-0.022160375,-0.014111694,-0.029382247,0.01366417,-0.029455645,0.014817702,-0.024585467,-0.0019078159,0.01875702,-0.035875127,-0.003161687,-0.026453154,-0.030973269,0.0063082944,0.005630624,0.023455523,0.003771399,0.0016681376,-0.0012370634,-0.030954013,0.011816266,0.032416355,0.03732589,0.018030414,-0.03392379,-0.00825221,0.019391995,0.0072873645,0.00756714,-0.005478314,-0.034795426,0.010865847,-0.011368819,0.020037547,0.023581825,0.0068440945,-0.018168237,-0.031704634,-0.002211634,-0.046861824,0.029733192,0.006009656,-0.010665488,0.0078964075,-0.001651064,-0.0025902567,0.006808961,-0.018667197,-0.0012057852,-0.002087731,-0.0099921385,0.0044933036,-0.008994091,-0.0018913201,0.020434435,0.03109269,0.00094916637,-0.007827161,0.020045917,-0.010403123,-0.0015248932,0.0074604256,0.019815974,-0.0056241695,-0.007277628,-0.011570686,-0.0053772526,-0.013179146,-0.0015023965,0.018240876,-0.007783678,0.025025817,-0.015489644,0.02574065,-0.018019382,0.0022568733,-0.016899906,-0.03308431,0.01197342,0.0040068245,-0.021629447,0.015786713,-0.022315217,0.007978023,-0.0003433731,0.006751127,0.03402833,-0.001268746,0.0074585103,-0.008988216,-0.01018571,-0.03185474,-0.003347431,-0.0018431245,0.0012278939,-0.0044854823,-0.015431687,0.014661753,0.0054337825,-0.0023486724,0.045711406,-0.0034594892,0.0345268,0.002682551,0.014364295,0.0073311757,-0.010132823,0.021186331,-0.014526562,-0.015932402,0.0141224265,-0.0013257366,0.0051266723,-0.010077137,-0.023340715,0.022623898,-0.087428726,-0.009635631,0.0076841447,-0.04063666,0.007960697,-0.0012544576,-0.007016488,-0.03452567,-0.00031838065,0.014356039,-0.017613424,-0.012180523,0.0192302,0.014764209,0.021369137,0.006677248,0.005924676,-0.0030364264,0.003189156,-0.02547048,0.013557695,-0.008382299,0.04321808,0.027403302,-0.0014555273,0.0011015689,-0.0015288988,0.024928967,0.020485085,0.014237399,-0.036075886,-0.02935383,0.0015006112,0.021038512,0.0133809075,-0.013755229,0.029292818,-0.00963844,-0.004539197,-0.0058230464,-0.0015397053,0.026724668,-0.007928184,-0.019846687,0.020819262,0.020497175,0.010800567,0.013531605,-0.003357379,0.007577465,-0.037427373,-0.024492312,-0.0032730806,-0.032818154,0.003167652,-0.008808231,0.0026190341,0.019252896,-0.008875326,0.019677546,-0.011978575,-0.038606584,-0.019160118,0.020766148,0.0073681795,0.0106436135,-0.0044504353,0.04357928,-0.0074357865,0.008295268,0.001996349,0.0005264386,0.007683648,0.020100944,0.0062764315,-0.015702283,-6.3424695e-06,0.0070955395,-0.00058848877,-0.016280489,-0.00697873,-0.026692221,-0.12744772,-0.0032149167,0.006910529,0.010614386,0.014451479,-0.01088353,0.03073367,-0.012472972,0.0029492446,-0.013196884,-0.011818109,0.002343035,0.003374868,-0.0035916632,-0.008545845,-0.014180028,-0.016608797,-0.019681817,0.0153652625,-0.046546265,-0.03254355,0.0003196702,0.0060228347,0.001104989,-0.0048882533,0.0005056186,-0.0038624397,-0.007042292,0.0042154095,-0.028093282,-0.010407956,-0.11487278,-0.0098242005,0.0021822287,-0.016398072,0.009420492,0.0071411985,-0.00632679,0.006962818,-0.018917443,-0.016723333,-0.021148048,-0.0229757,-0.01580188,-0.011423309,0.01884644,0.10510206,0.0070713544,-0.016492652,-0.020505624,-0.020170456,-0.013991318,-0.023486491,-0.0071000913,0.039325416,0.013711025,-0.003056001,0.03963382,-0.0034844736,-0.028534792,0.00068831124,0.007162784,0.0062207608,-0.013731474,-0.0049838675,0.033089522,-0.00019833521,-0.00938204,-0.005311716,0.006259333,0.015918477,0.040188923,0.007737757,-0.0010252725,-0.015031664,-0.0011726321,-0.0052076895,0.008264055,-0.036879536,-0.01488143,0.011306056,-0.019284207,-0.044491973,0.010499632,-0.013465923,0.008520424,0.005862284,-0.015635611,-0.0043242658,0.021761747,-0.0006202121,0.014581716,0.01590792,-0.006298188,0.012536641,-0.018790957,0.0015466164,0.0037300664,0.044344693,0.016787501,-0.0021783325,-0.012493205,0.006276547,0.014345149,-0.0016520944,0.003917243,-0.025546301,-0.009480204,0.00927158,0.032121945,-0.012298522,0.009855737,-0.0023944464,-0.02111632,-0.029697461,0.005262815,0.010055279,0.021865927,0.009795636,0.0076726284,-0.002955416,0.02396653,0.010343141,-0.006469517,0.002295529,0.0071638627,0.023380052,-0.008105426,0.032530695,0.0020892865,-0.0066838716,0.03177315,-0.018543556,-0.0041990154,-0.024477337,0.023867834,0.019496871,0.004341828,0.028532956,0.042581618,0.018217066]",{"tags":37,"relatedLang":47,"relatedPosts":51},[38,40,42,44,45],{"name":14,"slug":39},"video-decoder",{"name":17,"slug":41},"psnr",{"name":13,"slug":43},"video-generation",{"name":16,"slug":16},{"name":15,"slug":46},"reference-conditioning",{"id":27,"slug":48,"title":49,"language":50},"refdecoder-reference-conditioned-video-decoder-en","RefDecoder adds reference conditioning to video decoders","en",[52,58,64,70,76,82],{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":26},"6ca303f0-7bd4-4bb2-be58-70d80da5ec40","why-ai-safety-teams-are-wrong-blame-only-alignment-zh","為什麼 AI 安全團隊錯把問題全怪在對齊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778947417022-ak55.png","2026-05-16T16:03:16.319335+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":26},"50b2e74e-7248-43a3-8775-451bf2569f33","why-fine-tuning-llms-domain-tasks-right-default-zh","為什麼針對領域任務微調 LLM 才是預設選項","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778916229431-9olk.png","2026-05-16T07:23:32.255569+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":26},"b9516feb-41d5-42a3-887e-7b47c5c9ffb7","atlas-one-token-visual-reasoning-zh","ATLAS 用一個 token 做視覺推理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032775-hp0w.png","2026-05-16T06:13:34.693651+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":26},"bfd03801-a200-4222-9370-8b441be41483","entitybench-long-range-video-consistency-zh","EntityBench 盯住長片一致性","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911845686-4mc8.png","2026-05-16T06:10:27.85068+00:00",{"id":77,"slug":78,"title":79,"cover_image":80,"image_url":80,"created_at":81,"category":26},"667b72b6-e821-4d68-80a1-e03340bc85f1","turboquant-seo-shift-small-sites-zh","TurboQuant 與小站 SEO 變化","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778840440690-kcw9.png","2026-05-15T10:20:27.319472+00:00",{"id":83,"slug":84,"title":85,"cover_image":86,"image_url":86,"created_at":87,"category":26},"381fb6c6-6da7-4444-831f-8c5eed8d685c","turboquant-vllm-comparison-fp8-kv-cache-zh","TurboQuant 與 FP8 實測結果","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778839867551-4v9g.png","2026-05-15T10:10:36.034569+00:00",[89,94,99,104,109,114,119,124,129,134],{"id":90,"slug":91,"title":92,"created_at":93},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"9f50561b-aebd-46ba-94a8-363198aa7091","openclaw-agents-manipulated-self-sabotage-zh","OpenClaw Agent 會自己搞砸自己","2026-03-28T03:03:18.786425+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"11f22e92-7066-4978-a544-31f5f2156ec6","vega-learning-to-drive-with-natural-language-instructions-zh","Vega：使用自然語言指示進行自駕車控制","2026-03-28T14:54:04.847912+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"a4c7cfec-8d0e-4fec-93cf-1b9699a530b8","drive-my-way-en-zh","Drive My Way：個性化自駕車風格的實現","2026-03-28T14:54:26.207495+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"dec02f89-fd39-41ba-8e4d-11ede93a536d","training-knowledge-bases-with-writeback-rag-zh","用 WriteBack-RAG 強化知識庫提升檢索效能","2026-03-28T14:54:45.775606+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"3886be5c-a137-40cc-b9e2-0bf18430c002","packforcing-efficient-long-video-generation-method-zh","PackForcing：短影片訓練也能生成長影片","2026-03-28T14:55:02.688141+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"72b90667-d930-4cc9-8ced-aaa0f8968d44","pixelsmile-toward-fine-grained-facial-expression-editing-zh","PixelSmile：提升精細臉部表情編輯的新方法","2026-03-28T14:55:20.678181+00:00",{"id":135,"slug":136,"title":137,"created_at":138},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00"]