AISpot

v2.2.2 預設啟用 agent 影像生成

OpenCode releases版本更新開發工具產品與方案訂閱動態

v2.2.2 把 experimental_image_generate 設為預設可用:只要接入的模型提供商支援影像模型,agent 無需額外配置即可生成影像,並可通過 experimental.image_generate.model 指定首選模型。新會話介面點選當前 Git 分支可直接開啟該專案的 worktree 選擇器;時間線引用、終端引用與瀏覽器標註統一為帶可選評論的編輯器。桌面端還恢復經典問題卡片,併為 Go 用量限制錯誤補充重置時間與帳戶連結。

影像生成從需配置變為預設開啟,接上支援影像模型的提供商即可用;同時 worktree 入口與統一標註編輯器改善日常會話操作,對常用這套 agent 桌面端與 TUI 的人最直接。

原標題:v2.2.2
閱讀原文

評分52 / 58(平均 55,門檻 60)
狀態未入選

原文

Highlights Image Generation by Default : When a connected provider supports image models, agents can now use experimental_image_generate without additional configuration. Set its preferred model through experimental.image_generate.model . ( #60258 ) Faster Worktree Selection : On the new-session screen, click the current Git branch to open that project’s “Run session in” worktree picker directly. ( #60263 ) Unified Quotes and Annotations : Timeline quotes, terminal quotes, and browser annotations now share a compact optional-comment editor, while source highlights remain attached to local drafts. ( #60164 ) Core Improvements Enabled experimental_image_generate by default when an image-capable provider is connected. ( #60258 ) Recorded the start time of each session’s latest execution. Bug fixes Rejected persistent terminal requests whose working directory no longer exists. Preserved media when sessions are forked or imported. Associated image-generation notices with the correct background job. TUI Bug fixes Preserved captured attachments and their original source paths. Thanks @OpeOginni . ( #60222 ) Prevented middle-truncated paths from overflowing at very narrow terminal widths. Improved Go usage-limit messages with the reset time and account page link. Correctly typed and displayed Go and ChatGPT usage-limit errors. Desktop Improvements Made the Git branch on the new-session screen open the current project’s worktree picker directly. ( #60263 ) Unified timeline quotes, terminal quotes, and browser annotations around one optional-comment workflow. ( #60164 ) Restored classic question cards by default and added Settings → Experimental → New question UI for the redesigned keyboard-oriented dock. ( #59723 ) Added a side-panel tab menu and kept close buttons visible when tabs overflow. Moved user-message actions into a compact menu. Redesigned the active todo pill to show the current task and concurrent work more clearly. Added consistent styling for searchable dialogs and composer pills. Bug fixes Stopped queued prompts from dropping frames as their content grows. Counted queued quotes and browser annotations as attachments. Reported failures to persist composer attachments through the normal error UI. Made Markdown fragment links scroll to the referenced heading. Kept each tab’s timeline state when switching directories. Ordered the inbox’s Working shelf by execution start time. Fixed side-panel launchers and automatic closing when the final tab is removed. Focused the matching navigation arrow while reviewing changes. Preserved pending model labels only when a queued prompt actually changes models. Added reset details, an account link, and a composer tray for Go usage limits. SDK Improvements Added if and until inputs for client.experimental.schedule.create , plus schedule check results and output fields. ( #60137 ) Exposed SessionInfo.time.busy for the latest execution’s start time. Added optional source metadata to prompt file attachments. Extensions Improvements Added optional title and subtitle fields to composer notes, source-relative quote positioning, and focus options for composer integrations. ( #60164 ) Bug fixes Corrected visible flowchart edges connected through Merman subgraphs.

相關報導

OpenCode releases● 精選AI 評分78

OpenCode 2.2.1 發布:新增 /goal 持續目標與影像生成

OpenCode 2.2.1 發布,最值得注意的是新增 /goal 命令:在 TUI 或 Desktop 輸入 /goal <objective>,會話會持續推進直到審查者確認目標達成,並可用 pause、resume、clear 控制。同時加入需手動開啟的 image_generate 工具,可建立或編輯圖片並指定預設模型;Desktop 預設啟用可搜尋的新模型選擇器,活動欄能直接檢視並停止後臺 shell 輸出。

X @Alibaba_Qwen● 精選AI 評分79

阿里開源 Qwen-Image-2.1-Turbo:8 步去噪出圖與編輯

阿里發布 Qwen-Image-2.1-Turbo,這是基於 Qwen-Image-2.1 的加速 checkpoint,只需 8 步去噪即可生成和編輯影像,權重已在 ModelScope 與 Hugging Face 開放。它沿用同一 7B 視覺生成架構,官方稱步數減少不犧牲品質,仍可從文字生成 2K 影像,並支援用自然語言繼續編輯,例如新增配飾或更換場景。Diffusers 中載入 QwenImage21Pipeline 即可使用推薦的 8 步取樣;

TechCrunch AIAI 評分67

Canva 發布首個自研影像生成模型 Canva Craft

Canva 推出首個自研影像生成模型 Canva Craft,所有使用者可在 Dream Lab 標籤頁使用,也可通過 Canva API 呼叫,並計劃經 MCP 接入 ChatGPT、Gemini 等第三方聊天機器人。模型分兩階段生成:先解析提示詞並規劃元素空間關係、相對大小、視覺層級與文字資訊,再輸出最高 2K、約 30 種寬高比的影像。它以 Canva 模板等設計圖而非照片訓練,文字更清晰、支援區域性精確編輯,還能用 Magic 轉成可編輯的分層設計;

量子位● 精選AI 評分87

Google 發布 Nano Banana 2.1 生圖模型,4K 直出並降價

Google 發布新一代生圖模型 Nano Banana 2.1,可 4K 直出,並升級蒙版編輯、背景替換與主體一致性。據 Arena 評測,其文生圖、多圖編輯與單圖編輯三項功能平均提升約 61.7 分,文生圖排名第四,僅次於 GPT。API 價格上,1K、2K 圖片生成價格減半,4K 降價約 25%;已上線 Gemini App、Google AI Studio,每天贈送 50 積分,開發者可用 gemini-nano-banana-2.1 接入付費 API。

機器之心● 精選AI 評分78

8B 智慧體自主呼叫工具做影片生成,VABench 從 56.5 升到 86.1

VideoGen-Agent 用 Qwen3-VL-8B-Instruct 做基座,在多輪互動中呼叫檢索、生成與驗證工具,先以 16K 教師軌跡監督微調,再用 8K 提示做多工強化學習。在 600 條提示的 VABench 上,Toolset 1 配置得 75.6 分,比基礎生成器 56.5 分高出 19.1 分;智慧體策略不變、僅升級生成工具後達 86.1 分,人類評估中 84.3% 的判斷偏好該配置。