[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-vision-transformer":3},{"tag":4,"articles":10},{"id":5,"name":6,"slug":7,"article_count":8,"description_zh":9,"description_en":9},"930555cf-8311-4236-ab25-773e09ee286c","Vision Transformer","vision-transformer",1,null,[11,20],{"id":12,"slug":13,"title":14,"summary":15,"category":16,"image_url":17,"cover_image":17,"language":18,"created_at":19},"92510ea0-af04-4078-9e4f-ac4365a55e48","outlier-tokens-diffusion-transformers-dsr-zh","DiT 的異常 token 怎麼救","這篇論文指出，RAE-DiT 管線裡的異常 token 會干擾注意力並造成影像瑕疵，並提出 Dual-Stage Registers 來同時處理編碼器與去噪器的問題。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778134864544-t1yt.png","zh","2026-05-07T06:20:31.06396+00:00",{"id":21,"slug":22,"title":23,"summary":24,"category":16,"image_url":25,"cover_image":25,"language":18,"created_at":26},"99ae4e8c-3fb2-4d46-8965-458bd16372c8","steerable-vi-vision-representations-zh","可被文字操控的 ViT 特徵","這篇論文把文字提示塞進 ViT 編碼過程中，讓影像特徵能朝指定物件偏移，同時盡量保留通用視覺用途。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775196412799-wufk.png","2026-04-03T06:06:34.443487+00:00"]