[{"data":1,"prerenderedAt":53},["Reactive",2],{"news-story-news-1792311983093-2dtwu":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":30,"category":37,"impact_score":38,"views_count":39,"likes_count":40,"related_entities":41,"tags":45,"image_url":52},"news-1792311983093-2dtwu","阿里淘天公布日产百万级多智能体生成架构 CommerceGen-Studio：6B SKU-DiT 击败 20B 开源模型与 Nano Banana Pro，微软提出测试时多样性对齐 MSR","Alibaba Unveils Million-Image/Day Multi-Agent CommerceGen-Studio (6B SKU-DiT Beats 20B Models and Nano Banana Pro) Alongside Microsoft's Test-Time Diversity Alignment (MSR)","阿里巴巴淘天集团发表生产级多智能体商业视觉生成系统 CommerceGen-Studio（arXiv:2610.20603，HF 52 赞），通过 PERL-Agent 闭环自进化与 6B 参数 SKU-DiT（RDR 三阶段强化蒸馏），支撑日均数百万张秒级生成，并在真实流量基准 CommerceGen-Bench 上以 9.43/9.45 分击败 20B Qwen-Image-Edit 及闭源 Nano Banana Pro；同期微软研究院提出 Mine, Sample, Render（MSR，arXiv:2610.21185），在测试时外置统计采样彻底解决大模型生成同质化（Mode Collapse）。","Alibaba Taobao & Tmall Group detailed CommerceGen-Studio (arXiv:2610.20603, 52 HF upvotes), a production multi-agent generation system serving millions of images daily within seconds. Its 6B SKU-DiT trained via Reinforce-Distill-Reinforce (RDR) scores 9.43 (single-image) and 9.45 (multi-image) on CommerceGen-Bench, outperforming 20B open models and closed-source Nano Banana Pro. Meanwhile, Microsoft Research introduced Mine, Sample, Render (MSR, arXiv:2610.21185) for training-free test-time diversity alignment.",[10,11,12,13],"阿里淘天 CommerceGen-Studio（arXiv:2610.20603）在生产环境支撑日均数百万张商品视觉生成、单图延迟仅数秒","通过 Reinforce-Distill-Reinforce（RDR）训练的 6B 参数 SKU-DiT 在 CommerceGen-Bench 单图/多图任务上斩获 9.43 / 9.45 分，越级击败 20B Qwen-Image-Edit（8.67 / 7.48）与闭源 Nano Banana Pro（9.39 / 9.33）","配合 PERL-Agent（Prompt-Eval-Refine-Loop）自进化闭环与 Layout Thinker Optimization 后，综合得分进一步提升至 9.62 / 9.48","微软研究院提出 Mine, Sample, Render（MSR，arXiv:2610.21185），通过从真实语料挖掘分类属性并在测试时显式采样，无需训练即可消除网页 UI 设计、智能体 3D 资产合成等五大场景的模式坍塌",[15,16,17,18],"Alibaba's CommerceGen-Studio (arXiv:2610.20603) serves millions of commercial product images per day in production with a latency of a few seconds per image","Its 6B SKU-DiT trained with Reinforce-Distill-Reinforce (RDR) scores 9.43 (single-image) and 9.45 (multi-image) on CommerceGen-Bench, beating 20B open models and closed-source Nano Banana Pro (9.39 / 9.33)","Combining SKU-DiT with the self-evolving PERL-Agent loop and Layout Thinker Optimization pushes scores to 9.62 (single-image) and 9.48 (multi-image)","Microsoft Research's Mine, Sample, Render (MSR, arXiv:2610.21185) externalizes test-time diversity via corpus-mined attribute sampling across webpage UI design, agent-generated 3D assets, user simulation, and text-to-image","### 核心背景与行业痛点\n在电商广告、网页 UI 设计与 3D 场景合成等工业级生成场景中，通用扩散模型与多模态大模型面临两大核心落地障碍：第一，**商业严格保真与低延迟的双重约束**——电商商品图要求 SKU 细节（Logo、文字、材质结构）零畸变，同时需要具备复杂营销排版推理能力，而 20B+ 级重型图像编辑模型不仅推理耗时漫长，且在多商品组合场景下极易发生属性串色；第二，**后训练对齐引发的生成同质化（Mode Collapse）**——经过 RLHF/DPO 对齐的生成模型在反复采样时往往收敛到单一审美模板（如千篇一律的网页布局或相似的商品背景）。阿里巴巴淘天集团提出的 **`CommerceGen-Studio`（`arXiv:2610.20603`）** 与微软研究院（Microsoft Research）提出的 **`Mine, Sample, Render (MSR)`（`arXiv:2610.21185`）** 分别从生产级多智能体协同与测试时外置多样性控制给出了高价值工程范式。\n\n### 架构亮点与底层机制\n1. **阿里 `CommerceGen-Studio`：三智能体协同与 `6B SKU-DiT` 强化蒸馏闭环（`arXiv:2610.20603`）**：\n   - **PERL-Agent（Prompt-Eval-Refine-Loop）**：将意图规划、质量评估与反思精修统一为多轮自进化智能体闭环，无需人工微调提示词即可迭代收敛至高转化率视觉构图；\n   - **Layout Thinker Optimization (LTO)**：针对复杂商品营销海报，联合执行空间布局规划与文案-视觉结构推理；\n   - **6B `SKU-DiT` 与 `Reinforce-Distill-Reinforce (RDR)` 范式**：为兼顾像素级商品保真与秒级响应，团队设计了仅 **6B 参数**的 `SKU-DiT`，并首创 **强化学习 → 步数蒸馏 → 二次强化学习（RDR）** 三阶段训练管线，在极少去噪步数下恢复高频纹理与结构一致性，支撑生产环境**日均数百万张图像、单图数秒级延迟**的高并发服务。\n2. **微软 `Mine, Sample, Render (MSR)`：测试时外置概率采样打破模式坍塌（`arXiv:2610.21185`）**：\n   - `MSR` 指出大模型不擅长作为“隐式随机数发生器”，但极擅长作为“受控条件渲染器”。框架分为三步：① **Mine**：从真实领域语料库中自动挖掘可解释的分类属性 Schema 及其经验联合分布；② **Sample**：在测试时根据真实分布对属性组合进行概率采样；③ **Render**：将采样到的细粒度属性作为显式条件交由生成模型渲染，无需任何重训练即可恢复真实世界长尾多样性。\n\n### 权威 Benchmark 与实测跑分对比\n在阿里巴巴基于真实线上流量构建的 **`CommerceGen-Bench`**（涵盖单图背景生成与多图商品组合两大核心任务，满分 10 分）上：\n- **仅 6B 参数的 `SKU-DiT`（无 Prompt 增强）**：斩获 **9.43 分（单图）** 与 **9.45 分（多图）**；\n- **全面超越 20B 级开源大模型**：显著领先 `Qwen-Image-2.1`（`8.95 / 8.90`）、20B `FireRed-Image-Edit-1.0`（`8.84 / 8.03`）、20B `Qwen-Image-Edit`（`8.67 / 7.48`）与 12B `FLUX.1 Kontext`（`8.76 / 6.72`）；\n- **击败顶尖闭源商业 API**：在无增强下即超越 `Nano Banana Pro`（`9.39 / 9.33`）、`Qwen-Image-2-Pro`（`9.11 / 9.34`）与 `Seedream 4.5`（`9.24 / 9.15`）；配合 `PERL-Agent` 自进化闭环后，综合得分进一步推高至 **9.62（单图）/ 9.48（多图）**。\n- **微软 `MSR` 跨五大学科验证**：在**前端网页 UI 设计生成、智能体驱动 3D 资产合成、用户行为仿真、气象报告生成与文生图**五大任务上，显著降低与真实语料分布的统计散度并拓宽覆盖率。\n\n### 开发者实战落地与开箱指南\n- **小参数 DiT 生产加速借鉴**：在垂直领域落地图像生成时，不必盲目堆砌 20B+ 参数；采用 `CommerceGen-Studio` 的 **RDR（RL → 蒸馏 → RL）** 管线训练 6B 级专用 DiT，并外挂多模态评测 Agent 组成 `Prompt-Eval-Refine` 闭环，是兼顾吞吐成本与 SOTA 质量的最佳路径。\n- **消除 Coding Agent 前端生成同质化**：使用 Coding Agent 生成前端网页或 3D 场景时，可直接引入微软 `MSR` 思路由脚本先从优秀设计库抽样配色、排版网格、字体层级与动效属性 JSON，再注入 Prompt 交由 LLM 渲染代码，立竿见影消除“AI 味”千篇一律模板。","### 核心背景与行业痛点\nProduction generative pipelines in e-commerce visuals, web UI synthesis, and 3D scene generation confront two major hurdles: first, **strict identity preservation under low-latency serving constraints**—commercial product imagery requires zero distortion on SKU logos and textures alongside complex spatial layout reasoning, where 20B+ image editing models suffer from high latency and multi-object attribute bleeding; second, **post-training mode collapse**—RLHF- and DPO-aligned models repeatedly converge onto a narrow set of stereotypical outputs when prompted for open-ended designs. Alibaba Taobao & Tmall Group's **`CommerceGen-Studio` (`arXiv:2610.20603`)** and Microsoft Research's **`Mine, Sample, Render (MSR)` (`arXiv:2610.21185`)** address these challenges through production multi-agent orchestration and test-time diversity alignment.\n\n### 架构亮点与底层机制\n1. **Alibaba `CommerceGen-Studio`: Multi-Agent Orchestration & 6B `SKU-DiT` via RDR (`arXiv:2610.20603`)**:\n   - **PERL-Agent (Prompt-Eval-Refine-Loop)**: Unifies prompt planning, automated quality evaluation, and iterative reflection into a closed-loop self-evolving agent workflow;\n   - **Layout Thinker Optimization (LTO)**: Performs joint layout planning and visual-textual structural reasoning for complex promotional compositions;\n   - **6B `SKU-DiT` with `Reinforce-Distill-Reinforce (RDR)`**: To achieve strict product fidelity at low latency, the team trained a compact **6B `SKU-DiT`** using a three-stage **Reinforce-Distill-Reinforce (RDR)** pipeline—applying RL alignment before and after step distillation—powering production traffic at **millions of images per day within a few seconds per image**.\n2. **Microsoft `Mine, Sample, Render (MSR)`: Externalizing Test-Time Diversity (`arXiv:2610.21185`)**:\n   - Recognizes that aligned generative models are poor implicit samplers but strong conditional renderers. `MSR` mines an interpretable categorical attribute schema and empirical joint statistics from a reference corpus (**Mine**), draws attribute configurations from those empirical statistics at inference time (**Sample**), and uses the LLM or diffusion model strictly as a conditional renderer (**Render**).\n\n### 权威 Benchmark 与实测跑分对比\nOn **`CommerceGen-Bench`** (built from live production e-commerce traffic across Single-Image and Multi-Image generation, scored out of 10):\n- **6B `SKU-DiT` (without prompt enhancement)**: Scores **9.43 (Single-Image)** and **9.45 (Multi-Image)**;\n- **Beats 20B Open-Source Models**: Outperforms `Qwen-Image-2.1` (`8.95 / 8.90`), 20B `FireRed-Image-Edit-1.0` (`8.84 / 8.03`), 20B `Qwen-Image-Edit` (`8.67 / 7.48`), and 12B `FLUX.1 Kontext` (`8.76 / 6.72`);\n- **Surpasses Frontier Closed-Source APIs**: Beats `Nano Banana Pro` (`9.39 / 9.33`), `Qwen-Image-2-Pro` (`9.11 / 9.34`), and `Seedream 4.5` (`9.24 / 9.15`), reaching **9.62 (Single-Image) / 9.48 (Multi-Image)** when paired with the full `PERL-Agent` loop.\n- **Microsoft `MSR` Across 5 Domains**: Validated across **webpage UI design, agent-generated 3D assets, user simulation, weather reports, and text-to-image generation**, closely matching empirical corpus distributions without retraining.\n\n### 开发者实战落地与开箱指南\n- **RDR Pipeline for Domain DiTs**: Instead of serving costly 20B+ diffusion transformers, teams can apply `Reinforce-Distill-Reinforce` to 6B-class models paired with an external `Prompt-Eval-Refine` agent loop to achieve superior fidelity at a fraction of the serving cost.\n- **Applying `MSR` to Coding Agent UI Generation**: When using Coding Agents to generate web frontends or 3D scenes, developers can eliminate repetitive \"AI template\" aesthetics by sampling layout, typography, and color schema tuples from an empirical design table before invoking the LLM as a conditional code renderer.",[22,25,27],{"title":23,"url":23,"type":24},"https://arxiv.org/abs/2610.20603","arxiv",{"title":26,"url":26,"type":24},"https://arxiv.org/abs/2610.21185",{"title":28,"url":28,"type":29},"https://huggingface.co/papers/2610.20603","docs",{"platform":31,"author_name":32,"author_handle":33,"author_avatar":34,"tweet_url":35,"published_at":36},"twitter","Alibaba Taobao & Tmall Group / Microsoft Research","@Alibaba_Qwen","https://unavatar.io/x/Alibaba_Qwen","","2026-10-18T08:26:23.093Z","Agent",91,41,0,{"models":42,"agents":43,"vendors":44},[],[],[],[46,47,48,49,50,51],"Alibaba","CommerceGen-Studio","Microsoft Research","Multi-Agent","Diffusion","Test-Time Scaling","https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20603.png",1792919798130]