Alibaba Qwen Releases Qwen-Image-2.1: A 7B Open-Weight Model for Image Generation and Editing
https://www.marktechpost.com/2026/09/21/alibaba-qwen-releases-qwen-image-2-1/📌 Qwen-Image-2.1：7B 就打贏對手的生圖模型
TL;DR：Alibaba 把生成與編輯合併進單一 7B checkpoint，靠 KV 快取重用衝出多圖編輯速度優勢。
模型越大越強，這個假設在 Alibaba Qwen 團隊手上被打了折扣。他們剛發布的 Qwen-Image-2.1，體積只有前代的三分之一左右，卻在自家 benchmark 上超越所有列出的開源模型，甚至贏過 Nano Banana 2.0。
🤔 從兩顆模型合併成一顆
去年（2025 年）8 月上線的初代 Qwen-Image 是 20B 參數、Apache 2.0 授權，文字轉圖和影像編輯分屬兩個獨立 checkpoint（編輯功能在 Qwen-Image-Edit）。Qwen-Image-2.1 把這兩件事收進同一個模型：一個 checkpoint 同時涵蓋文字轉圖、多參考圖編輯、局部編輯，以及透明 RGBA 輸出。視覺生成核心是 7B 參數、32 層 single-stream DiT，Qwen 團隊稱其為 Qwen-Image 系列中「最均衡、最具成本效益」的版本。
值得留意的是，這個 7B 只是 diffusion transformer 本身的規模；完整 pipeline 還要載入一顆 8B 的 Qwen3-VL 編碼器，做容量規劃時得把這塊算進去。
🧩 速度秘密在 attention mask 的分工
Qwen 把這套機制稱為 mixed-granularity attention：文字 token 使用 token-level causal mask，圖片 token 則在各自影像內使用 chunk-level bidirectional mask。條件前綴（condition prefix）被放在 noisy latent 之前，因此它永遠不會關注（attend to）noisy latent，這代表它的 key/value 在整個去噪過程中維持固定。
實際運作上，模型只在第一步計算一次文字與輸入圖片的表徵，之後每一個去噪步驟都重複使用同一份 prefix KV 快取。參考圖片數量越多，這份快取省下的重複計算就越多，這也是官方強調的多圖編輯速度優勢的來源。
📊 Qwen-Image-Bench 上的名次
在 Qwen 自家的 Qwen-Image-Bench 上，Qwen-Image-2.1 拿下 60.28 分，高於 Nano Banana 2.0 的 59.82，也高於榜上所有其他開源模型，包括 32B 的開源模型 FLUX 2 Max（55.33 分）。不過在封閉模型陣營中仍有 6 款分數更高，榜首是 GPT Image 2.5 Sunburst 的 67.01 分。
🧩 部署與生態系支援
環境需求為 PyTorch 2.4.0 以上、transformers 5.17 以上、從原始碼安裝的 Diffusers，以及 accelerate、pillow。同一個 pipeline 只要傳入 image= 參數（1 張或多張參考圖）即可切換到編輯模式；GPU 記憶體吃緊時可用 pipe.enable_model_cpu_offload() 降壓。
Day 0 支援涵蓋 Diffusers、ComfyUI、vLLM-Omni、SGLang、LightX2V。其中 vLLM-Omni 提供 FP8 量化、prefix KV 快取、CUDA Graph decode 與 tensor parallelism；SGLang 則有 Cache-DiT、CUDA graphs、多 GPU 平行化與元件卸載（component offload）；ComfyUI 已備妥原生節點與轉換後權重。硬體支援也不侷限 NVIDIA，還涵蓋 AMD Radeon（透過 ROCm）以及透過 FlagOS 支援的 8 種晶片平臺。
Qwen 團隊同時釋出兩個 prompt 改寫模型，是基於 Qwen3.5-VL 9B 微調出的文字轉圖、編輯專用 checkpoint，可將簡短 prompt 擴寫成詳細描述，並能自動挑選長寬比。
⚠️ 研究可用，商用要另外談
目前的權重與授權定位是研究與評估用途；若要商業部署，需要向 Qwen 另外取得授權，這點和初代 Apache 2.0 的開放程度不同，評估導入前務必先確認授權範圍。
🎯 實務啟示
對於需要文字轉圖與編輯共用一套 pipeline 的團隊，Qwen-Image-2.1 把部署複雜度從「兩個 checkpoint 各自維運」降成「一個模型走天下」，加上 prefix KV 快取對多參考圖編輯場景特別友善，值得在需要頻繁多圖迭代編輯的產品中優先評估；但商用授權與 8B 編碼器帶來的實際顯存需求，都是上線前該先算清楚的成本項。
🔗 來源
- 標題：Alibaba Qwen Releases Qwen-Image-2.1: A 7B Open-Weight Model for Image Generation and Editing
- 作者／機構：Michal Sutter, MarkTechPost
- 連結：https://www.marktechpost.com/2026/09/21/alibaba-qwen-releases-qwen-image-2-1/
#QwenImage #AlibabaAI #ImageGeneration #ImageEditing #DiffusionModel #OpenWeightModels #DiT #KVCache #Diffusers #AIInference
原始資料 MarkTechPost · 收集於 2026-09-21
摘要原文
Alibaba’s Qwen team has released Qwen-Image-2.1 , a unified text-to-image generation and image editing model. Its visual generation component has 7B parameters across 32 single-stream DiT layers. One checkpoint covers text-to-image, multi-reference editing, local edits, and transparent RGBA output. Is it deployable? Yes, for research and evaluation. Day 0 support covers Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V. Commercial deployment needs a separate license from Qwen. The original Qwen-Image shipped in August 2025 as a 20B model under Apache 2.0. Editing lived in a separate Qwen-Image-Edit checkpoint. Qwen-Image-2.1 folds both jobs into one model at about a third of the size. Qwen team calls it the most balanced and cost-effective model in the Qwen-Image series. One important thing to note here for capacity planning: the 7B figure covers the diffusion transformer only. The pipeline also loads an 8B Qwen3-VL encoder. The GitHub Repo lists 4 components: The attention mask is where the speed comes from. Text tokens use a token-level causal mask. Image tokens use a chunk-level bidirectional mask within each image. Qwen calls this mixed-granularity attention. The condition prefix sits before the noisy latent, so it never attends to it. Its keys and values therefore stay fixed across denoising steps. The model computes text and input images once, at the first step. It reuses that prefix KV cache for every remaining step. Savings grow with the number of reference images, which explains the multi-image speed claim. The research team compares models on Qwen-Image-Bench, Qwen’s in-house benchmark. On that chart, Qwen-Image-2.1 scores 60.28 overall. That places it above Nano Banana 2.0 at 59.82 and every listed open-weight model. FLUX 2 Max, a 32B open model, sits at 55.33. 6 closed models score higher, led by GPT Image 2.5 Sunburst at 67.01. Install PyTorch 2.4.0 or later, transformers 5.17 or later, Diffusers from source, accelerate, and pillow. Then: The same pipeline handles editing when you pass image= with 1 or more references. On smaller GPUs, pipe.enable_model_cpu_offload() reduces memory pressure. For serving, vLLM-Omni adds FP8 quantization, prefix KV caching, CUDA Graph decode, and tensor parallelism. SGLang adds Cache-DiT, CUDA graphs, multi-GPU parallelism, and component offload. ComfyUI ships native nodes and converted weights. Beyond NVIDIA, the release covers AMD Radeon GPUs via ROCm and 8 chip platforms via FlagOS . Qwen team also released 2 prompt-rewriting models, fine-tuned Qwen3.5-VL 9B checkpoints for text-to-image and editing . They expand short prompts into detailed ones and can pick an aspect ratio. Check out the Model Weights , GitHub Repo , and Technical Details . 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. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Alibaba Qwen Releases Qwen-Image-2.1: A 7B Open-Weight Model for Image Generation and Editing appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成