[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1ik1ksbyla5x6":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":12,"bodyZh":13,"bodyEn":14,"sourceName":15,"sourceUrl":16,"updatedAt":17},"bfl-flux-3-image","NEWS","2026-10-04T05:08:00.000Z","Black Forest Labs 发布 Flux 3 Image：边界框 + 10 张参考图 + 4K 多步局部编辑","Black Forest Labs releases Flux 3 Image: bounding-box + 10 reference images + 4K multi-step local editing","BFL 发布 Flux 3 Image：原生支持边界框精准构图、最多 10 张参考图、最高 4K 输出分辨率、多步局部编辑——把\"可控生成\"从概念拉到工作流层面。","BFL released Flux 3 Image: native bounding-box composition, up to 10 reference images, up to 4K output, and multi-step local editing — pulling controllable generation from concept to workflow.",[],[],"### 是什么\n\n10 月 3 日，Black Forest Labs（BFL）发布新一代图像生成模型 **Flux 3 Image**。相比 Flux 1.1，核心升级落在\"可控\"与\"工作流化\"：\n\n- **边界框（bounding box）原生支持**：在 prompt 之外精确指定主体位置与尺寸\n- **最多 10 张参考图**：单次生成可同时参考多张图，统一风格 \u002F 角色一致性 \u002F 视角\n- **最高 4K 输出分辨率**：直接出大图，不必后处理\n- **多步骤局部编辑**：在同一画布上多次局部修改而不破坏整体一致性\n\n### 为什么值得关注\n\n此前可控生成主要靠 ControlNet、IP-Adapter 等外挂模块，参数面分散在不同模型之间。Flux 3 把这些能力内建到基座里——对生产环境更友好：一条 prompt 路径走完构图、参考、分辨率、局部编辑，不再\"先 base、再 inpaint、再 upscale\"三段拼。\n\n### 与同类对比（图像生成模型）\n\n| 维度 | Flux 3 Image | Qwen-Image-2.1 | Stable Diffusion 3.5 |\n|---|---|---|---|\n| 边界框原生 | ✓ | 部分 | ✗（需 ControlNet） |\n| 多参考图 | 最多 10 | 1–2（外挂） | ✗ |\n| 输出分辨率 | 4K | 2K | 1K（外挂 upscale） |\n| 局部编辑 | 原生多步 | 部分 | inpaint 单独管线 |\n\n### 本地部署角度\n\nFlux 系列一贯支持 diffusers \u002F ComfyUI \u002F Forge；4K 输出与多参考对显存要求较高，建议 24GB+ 消费级或 16GB+ 工作站（量化档位待社区反馈）。","### What it is\n\nOn October 3, Black Forest Labs (BFL) released **Flux 3 Image**. The core upgrades over Flux 1.1 sit on controllability and workflow:\n\n- **Native bounding-box support**: precisely specify subject position and size beyond the prompt\n- **Up to 10 reference images**: single generation can take multiple references at once — unified style, character consistency, viewpoint\n- **Up to 4K output**: direct high-res output, no post-upscale\n- **Multi-step local editing**: repeated local edits on the same canvas without breaking global consistency\n\n### Why it matters\n\nControllable generation previously leaned on external modules (ControlNet, IP-Adapter, etc.) with parameters scattered across models. Flux 3 bakes them into the base — friendlier for production: one prompt path covers composition, reference, resolution and local editing, instead of the \"base → inpaint → upscale\" three-stage duct tape.\n\n### Comparison (image generation)\n\n| Dimension | Flux 3 Image | Qwen-Image-2.1 | Stable Diffusion 3.5 |\n|---|---|---|---|\n| Native bbox | ✓ | Partial | ✗ (needs ControlNet) |\n| Multi-reference | up to 10 | 1–2 (external) | ✗ |\n| Output resolution | 4K | 2K | 1K (external upscale) |\n| Local editing | Native multi-step | Partial | inpaint as separate pipeline |\n\n### Local deployment angle\n\nThe Flux series has long supported diffusers \u002F ComfyUI \u002F Forge. 4K output and multi-reference are VRAM-hungry — 24 GB+ consumer or 16 GB+ workstation recommended; quantization variants await community feedback.","AI 日报（ai6s.net）","https:\u002F\u002Fai6s.net\u002F6ac0377c05257b0857147348.html","2026-10-04T01:02:09.842Z"]