Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model
https://www.marktechpost.com/2026/10/09/alibaba-qwen-releases-qwen-image-2-1-turbo-an-8-step-7b-image-model/📌 Qwen-Image-2.1-Turbo：同一顆 7B 模型，8 步出圖取代 40 步
TL;DR：Alibaba Qwen 把 Qwen-Image-2.1 做成 8 步推論版本，架構不變、速度大幅提升，還附上便宜的 API 選項。
生成一張圖要跑 40 次 denoising，對互動式應用來說始終是個延遲負擔。Qwen 團隊這次直接把步數砍到 8 步，架構完全沒換。
🤔 同一顆模型，兩種速度
Qwen-Image-2.1-Turbo 是 Alibaba Qwen 團隊基於開源權重模型 Qwen-Image-2.1 做出的加速版 checkpoint。它用 8 個 denoising 步驟完成生成與編輯，相較基礎模型預設的 40 步,等於同一個 7B 架構下少跑 5 倍的步數。Turbo 版保留與基礎模型相同的 7B 視覺生成架構，可以直接用 Diffusers 的 QwenImage21Pipeline 載入。
🧩 靠 prefix KV caching 把條件運算只算一次
底層架構是單串流（single-stream）DiT，32 層、7B 參數，採用 block-causal attention：文字 token 用 token 層級的因果遮罩（causal mask），圖像則用 chunk 層級的雙向遮罩（bidirectional mask）。這個注意力設計正是 prefix caching 能成立的原因,輸入圖像與文字的條件資訊在第一步就算完，之後每一步都重複利用這份快取。因為只跑 8 步，被快取的 prefix 幾乎覆蓋了大部分的條件運算成本。Turbo checkpoint 內建了官方建議的 8 步取樣排程，生成預設使用 CFG=1。
模型卡展示涵蓋 8 大類場景，包括人像、人體姿態、透明圖層、排版與海報設計、UI 版面。編輯範例則有單張圖像轉換、多參考圖合成，以及 4 張圖的室內場景合成。基礎模型本身支援最多 10 張參考圖，以及用圈選、手繪標註或遮罩做局部編輯。
💻 怎麼跑起來
環境需要從原始碼安裝 Diffusers，搭配 transformers>=5.17.0。這個 checkpoint 相依 Diffusers PR #14950，該 PR 新增了 pipeline 內設定取樣 sigma 的功能。做編輯任務時，傳入 image=input_image 搭配指令式 prompt 即可。支援的解析度預設從 2048×2048 的正方形到 2752×1536 的 16:9 都有。
有個細節要特別注意：單獨設定 num_inference_steps 不會覆蓋 checkpoint 內建的取樣排程，只有明確傳入 sigmas 參數才會生效,而 Qwen 也特別註明其他排程組合未經測試。
📊 價格與速率限制
Alibaba Cloud Model Studio 同時上架了 Turbo 與 Pro 兩個版本。qwen-image-2.1-turbo 每張圖 CNY 0.1，速率上限 120 RPM；qwen-image-2.1-pro 每張圖 CNY 0.25，速率上限 20 RPM。換算下來 Turbo 每張圖便宜 2.5 倍，同時請求速率上限是 Pro 的 6 倍。
⚠️ 使用上的限制
官方僅驗證並保證內建的 8 步取樣排程,其他步數或自訂 sigma 組合並未經過測試，想要調整生成品質與速度的取捨空間時要留意這一點。
🎯 實務啟示
對需要把圖像生成放進互動流程（像是即時編輯預覽、批量素材生成）的團隊，Turbo 版的賣點不是新能力,而是用同一套架構把延遲和成本同時壓低。如果產線本來就是 Qwen-Image-2.1，切換到 Turbo checkpoint 幾乎是零架構成本的升級，只需確認 Diffusers 版本與 PR 相依是否到位。
🔗 來源
- 標題：Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/10/09/alibaba-qwen-releases-qwen-image-2-1-turbo-an-8-step-7b-image-model/
#QwenImage #Alibaba #DiffusionModel #ImageGeneration #OpenWeights #Diffusers #TextToImage #AIImageEditing #DiT #GenerativeAI
原始資料 MarkTechPost · 收集於 2026-10-09
摘要原文
Alibaba’s Qwen team has released Qwen-Image-2.1-Turbo , an accelerated checkpoint of its open-weight Qwen-Image-2.1 model. It generates and edits images in 8 denoising steps instead of the base model’s 40-step default . For developers, that means 5x fewer denoising steps on the same 7B architecture, plus a hosted API option. Qwen-Image-2.1-Turbo is an 8-step image generation and editing checkpoint from Alibaba’s Qwen team. It is built on Qwen-Image-2.1 . Turbo keeps the same 7B visual generation architecture. It loads directly with QwenImage21Pipeline in Diffusers. The checkpoint ships with its recommended 8-step sampling schedule saved inside. Generation uses CFG=1 by default. Prefix KV caching reuses the text and reference-image context across denoising steps. The underlying architecture is a single-stream DiT with 32 layers and 7B parameters. It uses block-causal attention. Text tokens get a token-level causal mask, while images use a chunk-level bidirectional mask. The attention design is what makes prefix caching work. Input images and text are computed once at the first step. Every later step reuses that cache. With only 8 steps, the cached prefix covers most of the conditioning cost. The Turbo model card showcase covers 8 categories. These include portraits, human poses, transparent images, typography and posters, and UI layouts. Editing examples include single-image transformation, multi-reference composition and 4-image interior composition. The base model supports up to 10 reference images and local edits via circles, painted annotations or masks. Setup requires Diffusers from source , along with transformers>=5.17.0 . The checkpoint needs Diffusers PR #14950 , which adds pipeline-configured sampling sigmas. For editing, pass image=input_image with an instruction prompt. Supported presets run from 2048×2048 square to 2752×1536 at 16:9. One thing to note here. Setting num_inference_steps alone does not override the saved schedule. Only an explicit sigmas argument does, and Qwen notes other schedules are untested. Alibaba Cloud Model Studio now hosts both Turbo and Pro. qwen-image-2.1-turbo costs CNY 0.1 per image in most regions, with a 120 RPM limit. qwen-image-2.1-pro costs CNY 0.25 per image, with 20 RPM. Turbo is 2.5x cheaper per image and allows 6x the request rate. Check out the ModelScope page , Hugging Face page , Pro API and Turbo API . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. The post Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成