v0.8.0

[ooolabdev]

OOOSplat 0.8.0

中文 | English

---

中文

OOOSplat 0.8.0 统一了多素材输入以及首次生成、补拍和继续生成链路，重点修复相机重建、透明素材、精确抽帧、任务恢复、日志进度和结果发布问题，并升级 MCP 与 WASD 漫游操作。

✨ 主要更新

统一多素材输入

• 输入区不再区分图片和视频，可同时添加或拖入多段 MP4／MOV、多个图片文件夹和散图集合。
• 每组素材独立检测类型、尺寸、帧率、Alpha 和 GPS 信息，并在同一个素材框中管理。
• 多段普通素材及完整 ERP 全景素材可以联合重建同一场景；无法建立空间联系的素材不会被强行合并。
• 每段视频独立保存抽帧设置和真实时间戳，不同帧率的视频不会被拼成同一条时间轴。
• 超过100张的图片集可直接在素材框中选择按顺序匹配或穷举匹配。
• 疑似素材类型、图片顺序和错误处理均整合到对应素材项中，不再集中到生成前弹窗。
• 暂不支持新导入 COLMAP 文件或文件夹；已有任务、已有可信重建和应用内部 COLMAP 流程不受影响。

统一生成、补拍与继续生成

• 首次生成、补拍和继续生成尽可能复用同一条处理流水线，根据已经存在且通过校验的成果决定下一步。
• 失败、取消和中断的任务统一使用“继续生成”，保留原质量档位、抽帧计划、执行路线和已经消耗的自动补救次数。
• 有效的画面准备、特征、匹配、重建、训练候选和发布结果可以继续复用。
• 已有可信训练候选时只恢复校验和发布，不重复启动 Brush。
• 补拍会冻结原项目快照，再在独立派生项目中执行，不修改原项目、原模型或已有编辑结果。
• 补拍支持一次添加多个新来源，并支持普通素材与完整 ERP 素材组合。
• 部分新增来源无法关联原场景时会逐项提示；全部新增素材均无法关联时安全停止。
• 完善旧普通项目和旧补拍任务的兼容校验，无法证明历史参数或坐标一致时不会冒险复用。
• 继续生成仍不是优化器断点续训；未完成的 Brush 训练会从保存的初始化重新开始。

相机重建与透明素材

• 同一普通图片文件夹默认共享一套相机内参，避免同一来源被错误拆分为大量独立相机。
• 普通图片集存在混合尺寸时会在素材输入阶段阻止生成，避免进入错误的相机解。
• 修复部分普通图片集相机初始化异常、内参不合理及注册率明显降低的问题。
• 低注册率且相机内参异常时，可在原几何阈值下检查有限的确定性初始化起点，无法获得有效结果时停止，不将错误重建送入训练。
• 普通透明图片为 SfM 单独生成 Alpha 预乘 RGB 和灰度 Mask，Brush 训练继续使用原始 RGBA 图片。
• 普通素材直接复用画面准备阶段生成的图片、Alpha 和相机数据，减少训练前重复转换。
• 精确抽帧使用直通时间戳并核对源帧映射，修复 FFmpeg 正常退出但实际少帧导致任务中止的问题。
• 普通素材快速／均衡／精细档的 Splat 上限统一采用有效像素与平均轨迹公式，系数分别为0.2／0.4／0.5。
• 全景继续使用独立预算，精细档按权威球面重建的稀疏点数×40规划上限，不因六面转换乘六。

任务进度、日志与发布恢复

• 补充图片准备、Alpha／Mask、焦距检查、缓存校验、训练图片准备、检查点保存和结果发布等应用日志。
• 特征提取恢复按真实画面数量显示进度，匹配和相机重建根据实际活动更新。
• 修复百分比在数值和“处理中”之间频繁切换、重建进度回退到0及总进度不实时更新的问题。
• 日志按任务、运行、操作和素材隔离，防止旧运行或迟到事件覆盖当前任务。
• 修复正常日志被错误显示为 [REDACTED]、已完成任务持续提示日志重新读取等问题。
• 发布过程增加可恢复事务，中断后可以继续校验、安装和登记结果，不重复训练。
• 只有模型校验、原子发布和任务结果保存全部成功后，才显示完成和最终 Splat 数量。
• 继续生成准备阶段的状态读取、文件写入、原子替换和缓存归档错误会保留具体操作及原始系统错误码。
• 修复前端等待旧运行、轮询泄漏及迟到终态导致的任务状态串用问题。

MCP 能力升级

• MCP Agent 可预检普通／全景和多组素材，获取需要用户回答的问题，再创建或继续任务。
• 支持查询逐素材抽帧、补帧、匹配、采用结果、实际训练配置和有界日志。
• 失败、取消和中断的任务可通过最新运行标识继续，避免迟到请求重复启动。
• 可在不打开预览的情况下导出已经保存的编辑结果。
• MCP 模型导出支持标准 PLY、压缩 PLY、SPZ 和 SOG。
• SPZ 坐标越界时仍须取得用户明确同意，才会调整本次导出副本。
• 保留路径授权、幂等请求、单任务运行锁和仅限本机回环访问的安全限制。

预览与界面交互

• 按住 WASD 时可自动锁定鼠标并直接转向，无需按住鼠标键。
• 松开最后一个 WASD 按键后恢复光标，按 Esc 可立即退出本次漫游。
• W/S 改为沿当前视线前后移动：抬头按 W 上升，低头按 W 下降；A/D 继续保持水平侧移。
• 鼠标锁定不可用或被拒绝时，保留原有 WASD 和鼠标拖动方式。
• 素材输入框、待确认问题、素材详情和已完成任务的只读状态采用统一布局。
• 新增日文 README；应用界面继续提供简体中文和英文。

可靠性与隐私

• 统计偏好读取异常时，本次会话停止匿名上报，不生成新的安装标识或覆盖原配置。
• 用户关闭匿名统计后，本次会话立即停止发送，保存失败不会重新打开。
• 补齐多素材、补拍、Bridge、失败阶段及实际 Brush 设备信息，同时避免不同运行之间串用数据。
• 诊断报告只使用当前失败操作对应的有界日志，并继续要求用户确认。
• 完善 Windows、macOS 和 Ubuntu 构建及 CI 检查。

💻 平台支持

• Windows 11 x64
• macOS 15+ Apple Silicon Alpha：M1 或更新芯片，未签名、未公证。
• Ubuntu 24.04 x86_64 Alpha

COLMAP 会根据平台及硬件条件使用 GPU 或 CPU。Brush 训练仍需要可用的图形处理后端，各功能的平台验证范围以项目文档为准。

📦 安装包

• Windows：OOOSplat-0.8.0-x64-windows.exe
• Apple Silicon macOS：OOOSplat-0.8.0-arm64-macos.dmg
• Ubuntu 24.04：OOOSplat-0.8.0-x64-linux.deb

---

English

OOOSplat 0.8.0 unifies multi-input handling with first-generation, reshoot, and continuation workflows. It focuses on camera reconstruction, transparent-media processing, exact frame extraction, task recovery, progress and logs, result publication, MCP capabilities, and WASD navigation.

✨ Highlights

Unified Multi-input Workflow

• The input area no longer separates images and videos. Add or drop multiple MP4/MOV files, image folders, and loose-image sets together.
• Each source is analyzed independently for type, dimensions, frame rate, Alpha, and GPS information, then managed in one input list.
• Multiple ordinary and full ERP panorama sources can be reconstructed as one scene. Inputs that cannot be spatially connected are not forcibly merged.
• Each video retains its own sampling settings and real timestamps instead of being joined into one timeline.
• Image sets above 100 images can select ordered or exhaustive matching directly inside the input area.
• Projection questions, image-order choices, and input errors are handled within the corresponding source item instead of a shared launch dialog.
• New COLMAP file or folder imports remain temporarily unavailable. Existing tasks, trusted reconstructions, and OOOSplat's internal COLMAP workflow are unaffected.

Unified Generation, Reshoot, and Continuation

• First generation, reshoots, and continuation reuse the same processing pipeline wherever possible and choose the next step from validated existing artifacts.
• Failed, cancelled, and interrupted tasks all use Continue generation while preserving the original quality, frame plan, execution route, and consumed automatic remedies.
• Valid prepared images, features, matches, reconstructions, training candidates, and publication results can be reused.
• A trusted training candidate resumes validation and publication without restarting Brush.
• Reshoots freeze the source project into an independent derived snapshot without modifying the original project, model, or saved edits.
• A reshoot can contain multiple new sources and combine ordinary inputs with full ERP inputs.
• Omitted sources are reported individually. Generation stops safely when none of the new inputs connect to the original scene.
• Legacy ordinary and reshoot projects receive compatibility validation before their saved artifacts are reused.
• Continue generation is still not optimizer-state resume. Incomplete Brush training restarts from its saved initialization.

Reconstruction and Transparent Media

• Ordinary images from the same folder share camera intrinsics by default instead of being split into unnecessary independent cameras.
• Mixed dimensions inside an ordinary image set are rejected during input analysis to avoid invalid camera solutions.
• Fixes camera initialization and invalid-intrinsic paths that could significantly lower image registration.
• When low registration and invalid intrinsics occur together, OOOSplat can check a limited set of deterministic initialization images under the original geometric thresholds.
• Ordinary transparent images use separate Alpha-premultiplied RGB images and grayscale masks for SfM, while Brush continues training from the original RGBA images.
• Ordinary inputs reuse prepared images, Alpha, and camera data instead of repeating conversion before training.
• Exact frame extraction now uses timestamp passthrough and validates source-frame mappings, fixing tasks that stopped after FFmpeg exited successfully with fewer frames.
• Ordinary Fast, Balanced, and Detailed Splat ceilings use the effective-pixel and mean-track-length formula with coefficients 0.2, 0.4, and 0.5.
• Panoramas retain a separate budget. Detailed uses 40× the authoritative spherical sparse-point count without multiplying by six.

Progress, Logs, and Publication Recovery

• Adds application activity logs for image preparation, Alpha/masks, focal checks, cache validation, training-image preparation, checkpoints, and publication.
• Feature extraction again reports real image counts, while matching and reconstruction update from actual activity.
• Fixes percentages switching between numeric values and “Processing,” reconstruction progress resetting to zero, and overall progress not updating.
• Logs are isolated by task, run, operation, and source so stale events cannot overwrite the current run.
• Fixes ordinary messages being displayed as [REDACTED] and completed tasks repeatedly reporting that logs are being reloaded.
• Publication now uses a recoverable transaction so interrupted validation, installation, and result registration can continue without retraining.
• Completion and the final Splat count are shown only after model validation, atomic publication, and task-result persistence succeed.
• State reads, file writes, atomic replacement, and cache-archive failures during continuation retain their operation context and original OS error code.
• Fixes frontend waiters accepting stale runs, leaking polls, or allowing late terminal events to overwrite the current task.

MCP Upgrades

• MCP agents can inspect ordinary, panorama, and multi-source inputs, obtain required user questions, and then create or continue tasks.
• Status includes bounded per-source sampling, backfill, matching, adoption, resolved training settings, and logs.
• Failed, cancelled, and interrupted tasks continue using the latest run ID, preventing stale requests from launching duplicate runs.
• Saved edits can be exported without opening preview.
• MCP model export supports standard PLY, compressed PLY, SPZ, and SOG.
• Out-of-range SPZ adjustment still requires explicit user approval before modifying the export copy.
• Path authorization, idempotent requests, the single-task lock, and loopback-only access remain enforced.

Preview and Interface Interaction

• Holding WASD can automatically lock the pointer for mouse-look without holding a mouse button.
• Releasing the last WASD key restores the cursor, while Esc exits the current navigation session.
• W/S now move along the full viewing direction: looking up and pressing W moves upward, while looking down moves downward. A/D remain horizontal.
• If Pointer Lock is unavailable or rejected, the existing WASD and mouse-drag controls remain available.
• Input lists, inline questions, source details, and read-only completed-task states now use a consistent layout.
• Adds a Japanese README. The application interface remains available in Simplified Chinese and English.

Reliability and Privacy

• If telemetry preferences cannot be read safely, anonymous reporting remains disabled for the session without replacing the existing installation identity or configuration.
• Turning anonymous telemetry off takes effect immediately for the current session, even if saving the preference fails.
• Completes multi-input, reshoot, Bridge, failure-stage, and actual Brush-device statistics without mixing data between runs.
• Diagnostic reports use only bounded logs belonging to the failed operation and still require user approval.
• Improves Windows, macOS, and Ubuntu builds and CI checks.

💻 Platforms

• Windows 11 x64
• macOS 15+ Apple Silicon Alpha: M1 or newer; unsigned and unnotarized.
• Ubuntu 24.04 x86_64 Alpha

COLMAP uses GPU or CPU according to platform and hardware support. Brush training still requires an available graphics backend. See the project documentation for feature-specific validation coverage.

📦 Packages

• Windows: OOOSplat-0.8.0-x64-windows.exe
• Apple Silicon macOS: OOOSplat-0.8.0-arm64-macos.dmg
• Ubuntu 24.04: OOOSplat-0.8.0-x64-linux.deb

---

This discussion was created from the release v0.8.0.