v1.6.6 — 固定角度与批量机位预设

批量机位现在支持固定角度与可复用预设，可直接按 45, -45, 90, -90 的顺序出图，无需每次重新调节。

• 新增「固定角度列表」，支持负数、小数与自定义角度；保持当前人物姿势、俯仰、焦距、构图和输出尺寸，批量后恢复原机位。
• 角度预设支持命名保存、同名更新、删除、JSON 导入和导出全部；导入同名预设自动编号，上次批量设置自动记住。
• 保留角度步进与收藏机位批量模式。预设保存在当前浏览器及访问地址下，跨设备请导入/导出 JSON。
• 包含局域网 HTTP 的 UUID 兼容修复，解决图片导入、读取连线等操作报 crypto.randomUUID is not a function。
• 更新中英文 README、操作截图及验证记录；无需新增模型。

**验证：**260 项 JavaScript 测试、130 项 Python 测试通过。Chrome 实际渲染了四个不同机位的引导图，并验证预设管理、刷新保留、窄屏布局和四个任务的提交顺序；队列回执使用本地测试接口，未重跑 Qwen 推理。验证记录。

**PR / Issues：**本次检查没有未关闭的 Issues。PR #7 的本机限制会让局域网上传、配置与姿势保存全部返回 403，已实测、回复并关闭，未合并该限制。

**更新：**在节点目录运行 git pull，重启 ComfyUI，关闭旧 Studio 后 Ctrl+F5 刷新。重新打开时标题应为 T8 · v1.6.6。

v1.5.7 — Canny and pose history fixes

修复 Canny 零阈值导致纯色区域全白，以及单击 IK / FK 控件但未拖动就清空重做的问题。阈值 0 继续可用，真实拖动的撤销/重做保持正常。

基于 v1.5.6 新做前端 10 项 + 独立子 Agent 后端 10 项联合检查，并追加 12 项原生浏览器交叉复核。226 项 JavaScript、103 项 Python 和 40 次隔离 HTTP 请求通过。检查范围与限制见 双语检查记录。

更新后重启 ComfyUI，关闭旧工作台并用 Ctrl+F5 刷新；标题应显示 T8 · v1.5.7。此前保存的错误引导图需重新生成并应用到节点。

---

Fix zero-threshold Canny filling constant regions with false edges, and native IK/FK no-motion clicks erasing redo. Zero thresholds remain valid; real drag undo/redo continues to work.

A fresh 20-scope parent/child audit against v1.5.6 plus 12 independent native-browser cross-checks passed. Validation includes 226 JavaScript tests, 103 Python tests and 40 isolated HTTP requests. See the bilingual audit for scope and limitations.

Restart ComfyUI, close the old studio and reload with Ctrl+F5. Confirm T8 · v1.5.7; regenerate and apply previously saved incorrect guides.

v1.5.6 — FK interaction and save validation fixes

Fix native FK gesture completion and rejected-save file publication.

• Finish bone rotation before undo, actor/bone switching, control edits, locking and mode changes; clean up native dragging on lost pointer capture or pointer cancellation. Preserve the incoming control value and the gesture's initial history entry.
• Normalize capture orientation, validate dimensions and serialize the complete scene before publishing PNG/snapshot files. Rejected dimensions or nonfinite metadata no longer leave orphan guides.
• Update both READMEs with the current actual ComfyUI screenshot and the 20-scope audit.

Validation: 224 JavaScript + 103 Python tests; 10 new native FK browser cases and existing 18 camera/IK cases; 63 new isolated backend HTTP requests; Intel D3D11 framing, pan, bookmark, batch and resource regression; actual local ComfyUI rotation/undo. Large-model generation and ONNX weights were not rerun.

更新后重启 ComfyUI，关闭旧 Studio 并 Ctrl+F5 刷新，确认标题 T8 · v1.5.6。Restart ComfyUI, close the old studio and reload with Ctrl+F5; confirm T8 · v1.5.6.

v1.5.5 · OpenPose and interaction fixes

修复原版 OpenPose 归一化坐标被当作像素导致骨架缩到左上角；身体、手、脸分别转换，保留原始关键点和缺失关节。

修复拖动中滚轮缩放被后续移动重置，以及撤销、切换/删除人物、键盘修改控件受到旧拖动干扰。原生 IK 手脚拖动会在撤销、切人、切模式或捕获中断前结束，避免误改已恢复的姿势或新人物。新控件设置和前一拖动可分别撤销。

联合检查 20 个范围：224 项 JavaScript、100 项 Python、18 项新原生浏览器交互和六项独立交叉复核全部通过。后端 80 个真实隔离 HTTP 请求通过；Intel D3D11 覆盖投影、引导图、收藏、批量及多人资源稳定性。本轮未重跑大型模型生成或 ONNX 权重推理。

详细检查记录 / Audit

更新后重启 ComfyUI，关闭旧工作台并 Ctrl+F5 刷新，确认 T8 · v1.5.5。旧版错误骨架需重新导入或提取，再应用到节点。

---

Fixes upstream normalized OpenPose coordinates being treated as pixels, which shrank the skeleton to the upper-left corner. Body, hands and face are converted independently while retaining raw keypoints and missing joints.

Wheel zoom survives continued dragging. Undo, role switching/deletion and keyboard controls finish the previous interaction, preserving separate undo entries. Native IK also ends before undo, role/mode switching or capture interruption so stale movement cannot modify a restored rig or another actor.

Fresh 20-scope joint audit: 224 JavaScript tests, 100 Python tests, 18 native browser cases and six independent cross-checks passed. Backend probes passed 80 real isolated HTTP requests; Intel D3D11 verification covered projection, guides, bookmarks, batches and resource stability. Large-model generation and ONNX weights were not rerun.

Restart ComfyUI and reload with Ctrl+F5 after updating. Confirm T8 · v1.5.5, then reimport/extract any affected old skeleton and apply the corrected guide.

v1.5.4 · Camera and pointer fixes

修复人偶/GLB 在侧面和背面视角下的平移方向；编辑、人物站位模式中键导航保持拍摄构图。单击、无操作、小数角度、取消拖动保留撤销/重做，修复其他指针干扰和鼠标捕获丢失后的误移动。

主 Agent 与独立子 Agent 完成 20 项新检查：217 JavaScript + 97 Python 全通过；真实显卡 32 组鼠标平移、36 组投影测试，收藏与批量恢复完全相同 PNG；128 个真实后端 HTTP 请求通过。双语 README、实际工作台截图和 完整检查记录 已更新。未重跑大型模型生成。

Fixed screen-aligned human/GLB panning at side/rear angles, free navigation in edit/position modes, undo/redo loss on unchanged or cancelled gestures, foreign-pointer interference and stray movement after lost capture.

A fresh 20-scope joint audit passed 217 JavaScript and 97 Python tests, 32 native GPU pan cases, 36 projection cases and 128 real backend HTTP requests. Bookmarks and batches restore exact guide PNGs. Bilingual documentation and an actual ComfyUI screenshot are updated. Large-model generation was not rerun.

更新后重启 ComfyUI，关闭旧工作台并 Ctrl+F5 刷新，确认 T8 · v1.5.4。
After updating, restart ComfyUI, close the old studio and reload with Ctrl+F5. Confirm T8 · v1.5.4.

v1.5.3 · 稳定性修复 / Stability fixes

基于 v1.5.2 完成新的 20 项父/子 Agent 联合检查，修复五类问题：

• 收藏缩略图失败时恢复相机、编辑模式、选中机位与撤销/重做。
• 浏览器姿势库过滤损坏条目；存储失败保留原列表，避免未保存条目或误删除。
• 错误上传格式与嵌套 multipart 返回明确 400，零资产写入。
• 非文本引导类型返回明确 400。
• 修复同一插件进程中 Windows 多实例并发写入的访问拒绝；其他磁盘错误保持原样。

验证：186 JavaScript + 97 Python，283 项通过；新增前端 255 组合；12 实例/12 线程/5 个全新目录共 480 次机位保存；真实 Intel D3D11 WebGL 与实际 ComfyUI 收藏、撤销、重做通过。双语 README、截图及 20 项检查报告 已更新。

更新后重启 ComfyUI，关闭旧工作台并 Ctrl+F5 刷新；标题应为 T8 · v1.5.3。本次无需新增模型。

---

A fresh 20-scope joint audit fixes bookmark rollback, browser pose-library recovery and persistence, malformed upload validation, invalid guide-type validation, and concurrent Windows writes within the plugin process.

Validation: 186 JavaScript and 97 Python tests; 255 new frontend combinations; 480 saves across 12 instances and 12 threads in five fresh directories; actual ComfyUI and Intel D3D11 rendering checks. Bilingual documentation and current UI screenshot are updated.

Restart ComfyUI and refresh with Ctrl+F5. The workbench should show T8 · v1.5.3. No new models are required. This round validates interface/rendering/storage contracts without rerunning large-model generation. Registry upload remains subject to independent review.

v1.5.2 — Editor stability and scene ZIP fixes

Failed edits could discard redo history, locked actors could still be resized, and replacing skinned GLBs left bone textures allocated. v1.5.2 restores failed transactions, enforces scale locks, releases GLB skeleton resources, and returns clear file errors for corrupt scene ZIPs.

• Preserve undo/redo and the original error when actor edits, cast layouts, pose copying, reconstruction, GLB import or mannequin loading fail.
• Keep scene scale controls disabled for locked actors; camera editing remains available.
• Dispose the imported GLB's skeleton when replacing it.
• Normalize corrupt scene ZIP failures into readable HTTP 400 responses.
• Refresh the bilingual README and actual ComfyUI workbench screenshot.

Validation: 178 JavaScript + 91 Python tests passed; a fresh joint 20-round audit, actual editor interactions, local ONNX, 7,680 manifest/prompt combinations, and hardware WebGL with 36 framing cases. The native 18-image CPU probe validates encoding contracts, not generation quality.

Full audit / 完整检查记录

本轮修复四类问题：失败操作丢失撤销/重做记录、锁定人物被全局尺度修改、切换蒙皮 GLB 后骨骼纹理未释放、损坏场景 ZIP 返回服务器错误。双语说明和真实 UI 截图已更新。

更新后重启 ComfyUI，关闭旧工作台并按 Ctrl+F5 刷新。标题应显示 T8 · v1.5.2。

v1.5.1 — Legacy scenes, OpenPose JSON and undo fixes

完成针对 1.5.0 的 20 轮联合检查，修复四个已复现问题。

• 修复旧场景 conditioning:null 保存时的 HTTP 500。
• 支持没有画布字段的标准像素坐标 OpenPose JSON；默认采用原图尺寸，无原图时采用输出尺寸，文件自带尺寸优先。
• 修复撤销后道具控件编辑旧对象，造成界面与保存位置/显示/锁定不一致。
• 修复精简旧场景缺少机位收藏等可选字段时无法打开，保留原姿势、朝向与尺度。

POSE 来源说明与真实输出同步，双语 README 更新实际截图和 20 轮检查记录。167 项 JavaScript、89 项 Python 测试通过；真实 ComfyUI HTTP、浏览器操作和 Intel 硬件 WebGL 复测通过。更新后重启 ComfyUI、Ctrl+F5，确认工作台 v1.5.1。

---

20 joint audit rounds against 1.5.0 fixed four reproduced defects: null-conditioning snapshot HTTP 500 errors, standard pixel-coordinate OpenPose JSON without canvas metadata, prop controls editing detached objects after undo, and sparse legacy snapshots failing to open. Legacy poses, headings and scales remain preserved.

POSE help now follows the actual guide source. Bilingual documentation includes a current workbench screenshot and the audit report. All 167 JavaScript and 89 Python tests passed, alongside real ComfyUI HTTP, browser and Intel hardware WebGL checks. Restart ComfyUI, reload with Ctrl+F5, and confirm v1.5.1 after updating. Existing model-generation limitations remain documented.

v1.5.0 · Multi-person scenes and identity references

多人工作台：同场编辑独立人物，统一输出引导图并自动绑定身份参考。

• 添加、复制、删除、排序、显示与锁定人物；独立动作、体型、站位、尺度、朝向、照片或文字身份。切换人物保持相机。
• 合影 DWPose 选人后复制姿势，支持 Fisher 多人骨架 PNG、OpenPose JSON 和 POSE_KEYPOINT 连线。
• 粗图、POSE、三维 Depth/Canny 共用机位；可选手指、可见关节和近白/远白深度。多人编码调用原生 Qwen Image 2.1，同步照片、提示词编号和实际生成画幅。
• 种子随机多人、全场模板、GLB 道具、一次握手对齐、场景 ZIP，以及冻结人物与照片的批量机位。
• 修复 V3 空 Autogrow 校验、静态原图误用旧人物数量、无原图的虚构图片编号、首次自定义提示词切换映射、模板排序身份和机位中心恢复问题。
• 修复三维 Canny 对比不足导致全黑，兼容双面 GLB；合法空图显示提示并继续输出，批量结束同步恢复提示状态。
• 双语 README 加入真实三联效果图、便携三人场景 ZIP、逐图验证记录。247 项回归通过，39 张原生 Qwen 实际生成并独立目视核验。主矩阵32张中7基本符合、2部分符合；九人及复杂遮挡仍不稳定。

照片姿势是二维估计；遮挡、隐藏肢体、模型的动作与身份遵循仍需检查。多人摆姿推荐 Qwen 底模，严格动作优先 POSE。更新后重启 ComfyUI、Ctrl+F5，确认工作台 v1.5.0 并重新应用。

---

Multi-person workbench: edit independent actors in one scene, render shared guides and bind identity references automatically.

• Add, clone, delete, reorder, hide and lock actors. Independent pose, shape, placement, scale, heading and photo/text identity. Actor switching preserves the camera.
• Copy selected people from DWPose group-photo detection; import Fisher multi-person skeleton PNG, OpenPose JSON or native POSE_KEYPOINT input.
• Coarse, POSE, scene Depth and Canny share the camera. Optional fingers, visible joints and depth polarity. The multi-person encoder delegates to native Qwen Image 2.1 and synchronizes references, prompt numbering and actual latent dimensions.
• Seeded cast randomization, scene templates, GLB props, one-time handshake alignment, portable scene ZIP and camera batches with frozen actors/photos.
• Fixed empty V3 Autogrow validation, stale cast counts for static photo guides, phantom image indices without an original, first custom-prompt mapping, template identity order and bookmark centers.
• Fixed blank low-contrast scene Canny, preserved double-sided GLBs and allowed intentional empty maps with a warning. Camera batches restore the original warning state.
• Bilingual README includes actual reference/guide/final comparisons, a portable three-person ZIP and per-image validation receipts. 247 regression tests pass; 39 actual native Qwen outputs were independently reviewed. In the 32-image main matrix, 7 broadly matched and 2 partially matched; nine people and complex occlusion remain unreliable.

Photo poses are 2D estimates. Occlusion, unseen limbs and generated pose/identity fidelity still require review. Use Qwen base for cast posing and prefer POSE for precise actions. Restart ComfyUI, Ctrl+F5, confirm v1.5.0 and Apply again.

v1.4.0 · Lens, photo poses and mannequin recovery

新增可调镜头透视、照片姿势复制、随机姿势与人偶资源恢复。

• 右侧「镜头透视」支持 12–200 mm 和 24 / 50 / 85 mm 预设；联动相机距离保持目标平面大小，预览、导出、收藏和批量共用镜头参数。默认原始镜头保持旧工作流。
• 左侧「从原图复制姿势到人偶」一次完成 DWPose 提取与重定向。默认保守平面复制可见方向；半身照片保留画外肢体的原姿势。全身可选推测立体并手动翻转深度。自动选择底模 + POSE。
• 姿势库新增混合、站姿、动作、坐姿随机变化，支持种子重现、撤销、继续编辑和保存。
• 人偶文件 / 贴图超时停止等待，加载失败显示明确错误和重试 / 修复入口。点击修复后才校验并补齐固定来源的人偶与贴图；不下载推理模型。
• 更新双语 README 和透视示意图，无新增模型依赖；照片复制复用 DWPose。

验证：100 项 JavaScript + 48 项 Python 测试通过，包含实际人偶骨骼 / 原生 IK、透视投影、快照与修复路由。未重新运行浏览器 WebGL 或 Qwen 推理。二维照片无法恢复准确三维深度，复杂遮挡和随机动作仍需检查。

更新：git pull，重启 ComfyUI，Ctrl+F5 强制刷新，确认标题 v1.4.0，再应用到节点。切换底模时连接 Studio 的 LoRA 强度输出。

---

Adjustable lens perspective, photo-to-mannequin pose copying, seeded random poses and asset recovery.

• Adjust 12–200 mm or choose 24 / 50 / 85 mm. Camera distance follows the lens to retain target-plane size. Preview, export, bookmarks and batches share the lens. Original lens preserves previous workflows.
• Copy photo pose to mannequin runs DWPose and retargeting in one action. Conservative planar mode copies visible directions, retaining unseen limbs in cropped photos. Full-body joints can use estimated depth and depth flips. Selects Qwen base + POSE.
• Random standing, action, seated or mixed poses support seed reproduction, undo, editing and saving.
• Mannequin and skin timeouts stop waiting. Clear errors offer Retry and explicit Repair, with SHA-256 checked bundled assets from pinned sources. No inference weights are downloaded by repair.
• Updated bilingual documentation and a lens illustration. No new model dependencies; photo copying uses DWPose.

Validation: 100 JavaScript and 48 Python tests passed, including actual mannequin bones / native IK, camera projection, snapshots and repair routes. Browser WebGL and Qwen inference were not rerun. 2D photos do not recover accurate depth; complex occlusion and random poses still need review.

Update with git pull, restart ComfyUI, reload with Ctrl+F5, confirm v1.4.0 and Apply to node. Connect the Studio LoRA strength output when switching to Qwen base.