FigFactory 1.0.18 · 单人姿态参考修正

修正单个人偶被生成为三视图和重复人物的问题：单人形体步骤只向模型提供该人偶的隔离渲染；补充单一视角、单一人物以及肩臂到手掌的解剖连续性约束。保留多人模式的完整构图参考。

FigFactory 1.0.17 · 本地形体真人化修正

修正本地分步生图的形体阶段：明确将三维人偶仅作为姿态与体型草图，输出单张真人摄影风格人物图；约束精确人数与单一视角，避免生成三视图、重复人物或直接复刻人偶外观。

FigFactory 1.0.16 · Noct-Q V3 本地生图

FigFactory 1.0.16 adds Noct-Q V3 Turbo as a second managed local image-generation stack. Its int8 safetensors checkpoint is downloaded from the pinned Civitai V3 Turbo file, with resumable transfer and SHA-256 verification; the int8 text encoder and shared VAE use Hugging Face with the selected mirror and automatic origin fallback. No model weights are bundled in regular installers. The Apple Silicon offline package remains limited to the previously verified Qwen Turbo stack and cannot download Noct-Q.

The previous Qwen-Image-2.1 Turbo 4-step stack remains available. Noct-Q uses its 6-step workflow without the Turbo LoRA. A New Image action clears cached prompt/reference inputs for a fresh request while preserving generation history and the 3D mannequin.

macOS arm64 and x64 DMGs are signed, notarized, stapled, and Gatekeeper-verified. Windows x64 installer is unsigned. Linux x64 installers are provided as DEB and RPM.

FigFactory 1.0.15 · ZenMux 姿态参考修复

FigFactory 1.0.15 improves ZenMux pose-reference handling. The cloud pose reference no longer contains numbered badges; people are matched by their head positions, and generation instructions clearly treat the mannequin as a pose and body-proportion guide that must be rendered as a natural human. Failed ZenMux requests are labeled as not producing a new image, with any image still on the canvas identified as historical.

Windows x64 installer is unsigned. Linux x64 installers are provided as DEB and RPM. macOS arm64 and x64 DMGs are signed, notarized, stapled, and Gatekeeper-verified.

FigFactory 1.0.14

FigFactory 1.0.14 adds staged local image generation. The app now generates the body first, edits clothing, replaces the scene, and applies the authoritative face reference last. Each stage uses the preceding image plus its own reference, can be skipped when its reference is absent or disabled, and saves its output in local history.

The scene stage now replaces the full background and lighting instead of carrying forward the studio backdrop. The body stage enforces the selected figure count and a single composition, avoiding repeated side-by-side people caused by mixed reference instructions. In local mode, explicit ZenMux prompt optimization focuses on the body and enabled pose reference; the result and errors are visible beside the controls.

Local tests passed: 28 unit tests, desktop UI tests, staged local UI tests, and real ComfyUI start/stop/restart/idle-shutdown tests. A real Apple Silicon scene test replaced a gray studio with the reference cafe while retaining the person. Results with other references still depend on the selected model and should be reviewed stage by stage.

Model weights are downloaded from Settings and are not bundled. macOS DMGs are signed and notarized. The Windows installer is unsigned.

FigFactory 1.0.13 · 本地 AI 生图与 ZenMux 提示词优化

提示词优化使用 ZenMux；生图服务可独立选择 ZenMux 或本地 Qwen-Image-2.1 Turbo。

• AI 设置支持下载和校验 Heretic W4A8、Q4_K_M 主模型、Turbo LoRA 与 VAE；安装包不含权重。
• 工作区和设置都有本地 AI 服务开关、运行状态与就绪提示；生成时按需启动，空闲两分钟自动关闭，关闭窗口也会停止服务。
• Apple Silicon 实机已通过 1024×1024 文生图、参考图编辑及服务启停测试。
• 提供 Apple Silicon 和 Intel 的签名公证 DMG、Windows x64 安装包，以及 Linux DEB/RPM。Intel Mac 可使用 ZenMux；本地托管引擎面向 Apple Silicon。Windows 安装包未进行代码签名。
• Landing Page 展示本地工作区与最新版下载入口。

验证：23 项单元测试、桌面工作流与本地/ZenMux 分流集成测试均通过；Mac DMG 在发布前验证公证票据和 Gatekeeper。

FigFactory 1.0.11

有权威人脸参考时：每次生图清除上一轮自动派生提示词与参考分析，重新捕获当前姿态，审视全部启用参考及当前最终提示词；保留当前手工文字，不复用分析缓存。无权威人脸参考时：保持原有流程，相同内容可复用已有参考分析，固定身份合同不生效。

每个人物只允许一张权威人脸图，放在输入图片前部，重复人脸会拦截生成。增加长期保存的固定身份合同，逐图写清借用内容、允许修改与不可变项。AI 检测到身份冲突时停止请求；不会自动将上一轮输出晋升为身份参考。修图也携带固定身份图，工作图不作为新身份锚点。

本次审视后的实际生图提示词可展开查看，支持 Markdown。准确名称 gpt-image-2 / openai/gpt-image-2 的 edits 请求省略 input_fidelity；ZenMux 2.5 型号保留各自现有配置。

验证：16 项单元测试、桌面集成回归（含有/无身份参考分支）、打包界面检查通过。测试采用模拟服务；实图身份一致性仍需审核。

FigFactory 1.0.9

添加或改为场景参考图时自动开启场景与照明，主动关闭后仍尊重开关。场景参考决定背景与照明，人偶截图只决定人物姿态，避免截图白底覆盖目标场景；原始场景和照明要求与综合提示词一起发送。

生图前检查参考类型、人物绑定、模型参考能力与数量，并显示本次启用/关闭状态。参考或文字变化时，即使所有提示词均有内容，也会由文字视觉模型核对全部启用参考，附加综合约束；手工文字保留，相同组合复用结果。

验证：15 项单元测试、桌面集成回归与打包界面检查通过。实际模型画面遵循程度取决于上游模型。

FigFactory 1.0.8

Before generation, use enabled references to fill empty scene, lighting, clothing and pose prompts while preserving manual text. Refresh automatic descriptions when references change, reuse descriptions for unchanged inputs, and stop if completion fails. For doodle edits with no instruction, compare marked and original images to infer an instruction; ambiguous marks require user text and never trigger image editing.

FigFactory 1.0.7

Add optional scene and lighting descriptions, streaming Markdown prompt generation, editable results and saved settings. Reference thumbnails now support face, clothing and scene roles; scene photos can be uploaded or pasted and inform text/image generation. Turning scene off excludes its references and invalidates final prompts optimized with the old scene.