Repository navigation
v0.6.0
Pre-releaseOOOSplat 0.6.0
中文
OOOSplat 0.6.0 加入 MCP、多任务管理、透明素材支持和更多导出方式,同时升级 COLMAP、训练监控与自动优化策略。
✨ 主要更新
MCP
现在可以使用 Codex 等本机 AI Agent 操作 OOOSplat:
- 创建并启动高斯泼溅生成任务。
- 查看任务列表、进度和结果。
- 增量读取运行日志。
- 取消正在运行的任务。
- Agent 断开连接后,已经启动的任务仍会继续运行。
- MCP 与手动生成共用同一套任务、进度和日志界面。
MCP 默认关闭,可在设置中启用。服务仅监听本机地址,并提供可以直接复制的 Codex 配置。
MCP 面向受信任的本机环境,不支持从公网直接访问。
多任务工作区
- 任务按照“新任务、已完成、未完成”分组显示。
- 可以创建多个任务草稿并调整顺序。
- 可选择在当前任务完成后自动运行下一个 GUI 草稿。
- MCP 创建的任务会标记为“由 AI Agent 创建”,但不会切换当前页面或触发自动续跑。
- 任务状态会持久保存,重新打开页面后可以继续查看。
- 修复恢复任务完成后从列表消失的问题。
- 生成过程中可以预览其他已完成项目。
- 已完成项目继续提供预览、补拍和导出功能。
透明图片与透明视频
改进透明 PNG、图片序列和透明 MOV 的处理:
- 保留素材原始颜色,不再对 RGB 进行 Alpha 预乘。
- 使用 Alpha 0.5 阈值生成严格的黑白 COLMAP 蒙版。
- Brush 继续使用原始图片,不改变输入颜色。
- 视频只解码和缩放一次,减少重复处理。
- Quality v2、恢复、桥接补帧和补拍使用相同规则。
- 旧版透明素材缓存会在需要时安全重建,不会删除已有项目或模型。
自动优化与训练
- 重新调整快速、均衡和精细档的增密参数。
- 快速档 Splat 上限为 500,000。
- 均衡档 Splat 上限为 1,000,000。
- 精细档会根据显存选择训练分辨率和 Splat 上限。
- High Emergency 的 Splat 上限提升至 1,000,000。
- Two-view tracks 只在开启自动优化且素材确实带有透明遮罩时启用。
- 精细档遇到明确的显存不足时,最多自动降低一次训练配置。
- 视频重建覆盖不足时仍可使用桥接补帧,并在失败时安全恢复原结果。
COLMAP 与 OOOBrush 升级
- Windows、Ubuntu 和 macOS 统一使用固定版本的 COLMAP 4.2.1。
- 升级并固定 OOOBrush 运行时。
- Windows 和 Linux 支持 CUDA、Caspar 与 Ceres。
- Caspar 不可用时会自动回退到 Ceres CPU。
- 改进 GPU 检测、显存不足判断和 Brush 选卡稳定性。
- 更清楚地区分显存不足、设备中断和普通运行错误。
进度、日志与错误诊断
任务详情现在可以显示:
- 当前阶段、迭代次数和完成进度。
- 实时总耗时与预计剩余时间。
- CPU、内存、GPU 和显存使用情况。
- Splat 数量、PSNR、SSIM 和实际训练配置。
- 实时日志与完整项目日志。
同时修复了总耗时不实时、日志关联错误任务和 Windows 路径显示异常等问题。
失败任务现在会显示更容易理解的错误摘要。用户也可以在确认脱敏内容后,主动发送错误报告。
预览与导出
- 新增 1920×1080 横屏视频导出。
- 保留 1080×1920 竖屏视频导出。
- 新增单文件离线 HTML 导出,无需联网即可查看模型。
- 离线查看器支持旋转、平移、缩放和动画播放。
- 导出保留当前模型调整、裁剪和编辑结果。
- 大型 PLY 使用流式加载,降低打开时的内存占用。
- 修复部分模型中 Gaussian 前后关系错误导致的显示问题。
🔒 本地计算与隐私
素材处理、COLMAP 重建、Gaussian Splat 训练和 MCP 任务均在本机运行。原始视频、图片和模型无需上传到云端。
匿名统计可以在“设置 → 隐私”中关闭。错误报告只有在用户查看脱敏预览并明确同意后才会发送。
💻 平台支持
- Windows 11 x64:支持兼容的 NVIDIA GPU,并可在需要时回退到 CPU。
- macOS 15+ Apple Silicon Alpha:支持 M1 及更新芯片,当前版本未签名、未公证。
- Ubuntu 24.04 x86_64 Alpha:内置固定版本的 COLMAP 和 Brush。
📦 下载
- Windows:
OOOSplat-0.6.0-x64-windows.exe - Apple Silicon macOS:
OOOSplat-0.6.0-arm64-macos.dmg - Ubuntu 24.04:
OOOSplat-0.6.0-x64-linux.deb
⚠️ 注意事项
- MCP 默认关闭,需要在设置中手动启用。
- MCP 免 token,仅建议在受信任的本机环境中使用。
- 自动优化仍属于实验性功能,可在设置中关闭。
- macOS 与 Ubuntu 版本目前为 Alpha。
- 离线 HTML 会包含完整模型,高精度项目可能生成较大的文件。
- 建议开始任务前预留足够磁盘空间,并更新显卡驱动。
English
OOOSplat 0.6.0 adds MCP, multi-task management, transparent-media support, and new export options. It also upgrades COLMAP, training monitoring, and automatic optimization.
✨ Highlights
MCP
OOOSplat can now be controlled by local AI agents such as Codex:
- Create and start Gaussian Splat generation tasks.
- View tasks, progress, and results.
- Read logs incrementally.
- Cancel active tasks.
- Accepted tasks continue running if the agent disconnects.
- MCP and manual generation share the same tasks, progress, and logs.
MCP is disabled by default and can be enabled in Settings. It listens only on localhost, requires no token, and provides a ready-to-copy Codex configuration.
MCP is intended for trusted local environments and is not exposed to the public internet.
Multi-task Workspace
- Tasks are grouped into New, Completed, and Unfinished sections.
- Multiple task drafts can be created and reordered.
- The GUI can optionally start the next draft after the current task finishes.
- MCP tasks are marked as created by an AI Agent, but do not change the current page or trigger automatic continuation.
- Task state is persisted and restored after reopening the application.
- Fixed resumed tasks disappearing after completion.
- Other completed projects can be previewed while generation is running.
- Completed projects retain Preview, Reshoot, and Export actions.
Transparent Images and Videos
Improved support for transparent PNG files, image sequences, and MOV videos:
- Original colors are preserved without alpha premultiplication.
- COLMAP masks use a strict black-and-white alpha threshold of 0.5.
- Brush continues to receive the original images.
- Transparent videos are decoded and scaled only once.
- Quality v2, resume, bridge recovery, and reshoot use the same rules.
- Older transparent-media caches are rebuilt safely without deleting projects or models.
Automatic Optimization and Training
- Rebalanced densification settings for Fast, Balanced, and High.
- Fast has a 500,000 Splat limit.
- Balanced has a 1,000,000 Splat limit.
- High selects its resolution and Splat limit according to available VRAM.
- The High Emergency limit has been increased to 1,000,000 Splats.
- Two-view tracks are used only when Auto Optimize is enabled and the source actually contains transparency.
- High can lower its training profile once after a confirmed out-of-memory failure.
- Video bridge recovery remains available and safely restores the original result if recovery fails.
COLMAP and OOOBrush Upgrades
- Windows, Ubuntu, and macOS now use pinned COLMAP 4.2.1 runtimes.
- The bundled OOOBrush runtime has been upgraded and pinned.
- Windows and Linux support CUDA, Caspar, and Ceres.
- Caspar automatically falls back to Ceres CPU when unavailable.
- Improved GPU detection, out-of-memory handling, and Brush adapter selection.
- Errors now distinguish GPU memory shortages, device interruptions, and general process failures more clearly.
Progress, Logs, and Diagnostics
Task details can now show:
- Current stage, iteration count, and progress.
- Live elapsed time and estimated remaining time.
- CPU, memory, GPU, and VRAM usage.
- Splat count, PSNR, SSIM, and effective training settings.
- Live logs and complete project logs.
This release also fixes inaccurate elapsed time, logs being associated with the wrong task, and Windows path-display issues.
Failed tasks now provide clearer error summaries. Users can also send an error report after reviewing and approving its redacted contents.
Preview and Export
- Added 1920×1080 landscape video export.
- Portrait export at 1080×1920 remains available.
- Added standalone offline HTML export.
- Offline HTML supports rotation, panning, zooming, and animation playback.
- Exports preserve model transforms, cropping, and editing results.
- Large PLY files now use streaming to reduce memory usage.
- Fixed incorrect Gaussian depth ordering on affected models.
🔒 Local Processing and Privacy
Media processing, COLMAP reconstruction, Gaussian Splat training, and MCP tasks run locally. Original videos, images, and models do not need to be uploaded to a cloud service.
Anonymous analytics can be disabled under Settings → Privacy. Error reports are only sent after the user reviews the redacted preview and explicitly agrees.
💻 Platform Support
- Windows 11 x64: Supports compatible NVIDIA GPUs and can fall back to CPU when required.
- macOS 15+ Apple Silicon Alpha: Supports M1 and newer Macs. The current build is unsigned and unnotarized.
- Ubuntu 24.04 x86_64 Alpha: Includes pinned COLMAP and Brush runtimes.
Windows, Ubuntu, and macOS builds have passed.
📦 Downloads
- Windows:
OOOSplat-0.6.0-x64-windows.exe - Apple Silicon macOS:
OOOSplat-0.6.0-arm64-macos.dmg - Ubuntu 24.04:
OOOSplat-0.6.0-x64-linux.deb
⚠️ Notes
- MCP is disabled by default and must be enabled in Settings.
- Token-free MCP is recommended only for trusted local environments.
- Auto Optimize remains experimental and can be disabled.
- macOS and Ubuntu releases remain Alpha.
- Offline HTML files include the complete model and may become large.
- Keep sufficient free disk space available and use an up-to-date graphics driver.
For complete documentation and usage instructions, see the project README.