Image and video圖像與影片
Preference arenas in September 2026 often put OpenAI's GPT Image 2.5 at the front of stills, and they split video between Google, Kuaishou, ByteDance, and MiniMax depending on the test. There is no single winner. There is a right tool for a poster, a product shot, and a ten-second scene.2026 年 9 月的偏好競技場，靜態圖常常把 OpenAI 的 GPT Image 2.5 放在前面；影片則依測試分給 Google、快手、字節跳動與 MiniMax。沒有單一贏家。海報、產品照與十秒場景，各有對的工具。
OpenAI · closedOpenAI · 封閉
The practical default when the prompt is a specification: layout, edits, a product held a certain way, words that must be spelled. September arena writeups put the Sunburst and Flare variants at the top of text-to-image preference. It lives inside ChatGPT and the API.提示是規格時的實用預設：版面、修改、以特定方式拿著的產品、必須拼對的字。九月競技場文章把 Sunburst 與 Flare 變體放在文字生成圖像偏好的前面。它住在 ChatGPT 與 API 裡。
Midjourney · closed · subscriptionMidjourney · 封閉 · 訂閱
Still the place to go when mood, material, and taste are the deliverable. Weaker when the image is a document full of words. No generous free tier. Use it for concept frames, then take the winner elsewhere for exact edits.氣氛、材質與品味就是交付物時，仍是該去的地方。圖像是滿是文字的文件時較弱。沒有大方的免費層。用來做概念畫面，精確修改再把贏家帶到別處。
Google · closedGoogle · 封閉
Reference-driven generation and conversational edits, with SynthID watermarks on the way out. Strong when you already work in Gemini and you need several images of the same person or product to agree.依參考生成，以及對話式修改，輸出帶 SynthID 浮水印。你已經在 Gemini 裡工作，又需要同一個人或產品的多張圖一致時很強。
Black Forest Labs · open weights on the dev/klein sideBlack Forest Labs · dev/klein 側是開放權重
The pipeline model. Photoreal stills, color control, and a ComfyUI graph you can repeat. This is the local answer on a 24 GB GPU or a 32 GB-and-up Mac. Hosted FLUX when you do not want the machine.管線模型。寫實靜態圖、色彩控制，以及可以重複的 ComfyUI 圖。這是 24 GB 顯示卡或 32 GB 以上 Mac 的本機答案。不想用機器時，用代管的 FLUX。
Ideogram · closed, some weights restrictedIdeogram · 封閉，部分權重受限
Posters, thumbnails, and layouts where the words are part of the picture. Proofread every glyph. Models still invent a letter when you look away.海報、縮圖，以及字是畫面一部分的版面。每個字形都要校對。你一看往別處，模型仍會發明一個字母。
Recraft · closedRecraft · 封閉
Logos, icons, and true SVG. Not a photoreal engine. Use it when the file has to land in a design tool as vectors.標誌、圖示，以及真正的 SVG。不是寫實引擎。檔案必須以向量進設計工具時用它。
ByteDance · Alibaba字節跳動 · 阿里巴巴
Strong on multilingual text inside the image and on editing-heavy production. Worth a trial for packaging, menus, and Chinese or mixed-language layouts where US models still slip.圖內多語文字，以及修改很多的製作，都很強。包裝、菜單，以及美國模型仍會滑掉的中文或混語版面，值得試。
xAI · inside GrokxAI · 在 Grok 裡
The fast path if you already pay for Grok and you want a still without changing apps. Compare it with GPT Image on one prompt before you make it the house style.你已經付 Grok，又想不換應用就出一張靜態圖時的快路。把它定成工作室風格之前，用同一個提示和 GPT Image 比一次。
Adobe · closedAdobe · 封閉
The choice when a client needs a commercial indemnity and the file must sit in Photoshop. Quality is not always the arena leader. The contract sometimes is.客戶需要商業保障、檔案又必須待在 Photoshop 時的選擇。品質不總是競技場領先。有時領先的是合約。
Google DeepMind · closedGoogle DeepMind · 封閉
The all-rounder when you want prompt adherence, native audio, and a clip you can extend. Higher price per second than the Chinese challengers. Use it for a hero shot, not for thirty drafts.你要貼近提示、原生聲音，以及可以延長的片段時的全能選擇。每秒價格高於中國的挑戰者。用在主鏡頭，不要用在三十個草稿。
Kuaishou 快手 · closed快手 · 封閉
Motion, camera moves, and value. A regular winner of blind votes and a sensible default for action, people, and social cuts. The 3.0 generation is the one to test in late 2026.動態、運鏡與價效。盲投的常勝軍，也是動作、人物與社群短片的合理預設。2026 年下半年該測的是 3.0 世代。
ByteDance · closed字節跳動 · 封閉
Image-to-video and multi-shot consistency. Start from a still you already like (from GPT Image, Midjourney, or FLUX) and let Seedance move it. Often the best "animate this frame" path.圖生影片與多鏡頭一致性。從你已經喜歡的靜態圖開始（GPT Image、Midjourney 或 FLUX），讓 Seedance 讓它動。常常是「讓這一格動起來」的最佳路徑。
MiniMax · closedMiniMax · 封閉
Fast drafts. Production writeups in September timed Hailuo among the quickest click-to-clip paths. Use the fast tier to find the shot, then re-roll the keeper on Kling, Seedance, or Veo.快速草稿。九月的製作文章把 Hailuo 計在點一下就出片的最快路徑之中。用快速層找鏡頭，再把要留的那條在 Kling、Seedance 或 Veo 上重跑。
Alibaba · Lightricks · open weights阿里巴巴 · Lightricks · 開放權重
The local and self-hosted video path, inside ComfyUI. Honest expectation: a short clip, a long wait, and a result you use as a draft unless the machine is a workstation. See the hardware page.ComfyUI 裡的本機與自架影片路徑。誠實的預期：短片段、長等待，結果當草稿，除非機器是工作站。見硬體頁。
Runway · closedRunway · 封閉
Director controls: camera paths, motion brushes, and a timeline a video person already understands. Choose it when the job is editing, not a single prompt.導演控制：鏡頭路徑、動態筆刷，以及影片工作者已經看得懂的時間軸。工作是剪輯而不是單一句提示時選它。
xAI · closedxAI · 封閉
Short clips next to the chat, useful for a quick motion test. Not the tool to build a commercial pipeline on until you have compared a keeper shot.聊天旁邊的短片，適合快速試動態。在你比過一條要留的鏡頭之前，不要拿它搭商業管線。
OpenAI's Sora was the famous name and is a poor thing to depend on without checking today's product page. Reporting through 2026 described consumer access being narrowed. If the API or app is still offered when you read this, test it. If it is not, do not architect around it.OpenAI 的 Sora 曾是響亮的名字，不先看今天的產品頁就不該依賴。2026 年的報導描述消費者使用權被收窄。你讀到這裡時若 API 或應用仍在，就測它。若不在，不要圍繞它設計。