v0.7.1

[ooolabdev]

OOOSplat 0.7.1

中文 | English

---

中文

OOOSplat 0.7.1 调整默认生成策略与精细档参数，升级补拍工作流，加入图片 GPS 辅助、抽帧比例控制和漫游调速，并改进桥接补帧、预览性能、导出及错误诊断。

✨ 主要更新

默认生成策略与自动重试

• 原“自动优化”的基础策略改为默认启用，保留自动补帧和闭环匹配。
• 三个档位均仅在实际图片存在透明区域时启用两视图三角化，全不透明的 Alpha 图片不启用。
• “自动优化”开关改为“自动重试”，默认开启，允许在符合条件的失败后降低点数上限或调整补拍方案。
• 明确 GPU 显存不足时，三个档位均可最多重试一次，点数上限取失败进程最后可信点数的80%，不降低训练分辨率。
• 完成详情仅在实际执行了自动重试后显示“已自动重试”。

缺少可信计数、无法实际降低上限或低于初始化点数时直接报错，不静默丢弃初始化点。关闭自动重试不会关闭正常补帧或后端兼容回退。

抽帧比例与桥接补帧

• 视频新增分级抽帧滑轨，支持手动调整和恢复默认；快速／均衡／精细默认基础抽帧目标保持6／8／12 FPS。
• 新任务使用自定义比例时，桥接补帧最多额外增加首次实际帧数的20%，总量不超过源视频可用帧数。例如首次360张，最多再补72张。
• 初始注册率低于80%、存在有效初始模型时，按未注册区间尝试一轮定向补帧，不重复抽取已有帧。
• 手动比例较高时也可从剩余源帧中补充；100%抽帧不再新增。首次选帧索引及最少30帧规则保持不变，源帧不足时全部采用。
• 未自定义比例时，仍使用快速／均衡／精细9／12／15 FPS候选预算。
• 全景补帧复用生成阶段已读取的真实时间戳，支持VFR和非零起始时间，不重复探测整段视频。
• 优化候选选择，避免为了补帧建立完整源帧数组；切换比例只重新规划，不重复检测素材。

1.2倍是帧数预算，不代表耗时只增加20%，也不保证重建成功。旧任务中断续跑保留原规则；失败后手动重试采用新策略。

精细档训练调整

• 精细档训练分辨率不再按显存分档，普通素材使用原始训练分辨率。
• 全景精细档保留原始 ERP，六面训练图片边长为 floor(ERP宽度 / 4)，不再限制为1024，不放大素材。
• 普通精细档按有效像素和平均重建轨迹长度计算 Splat 上限，系数为0.5；完全透明的像素不计入。
• 全景继续使用独立点数预算，实际参数以任务保存的配置为准，不按六面数量乘六。

快速和均衡维持现有分辨率策略。Splat 上限不是最终点数或画质保证；原始分辨率训练可能明显增加显存、内存和磁盘占用。

补拍工作流升级

• 普通补拍支持不同分辨率、宽高比及设备／镜头；全景项目支持完整 ERP 视频和图片补拍。
• 通过简单问答选择处理方式，并判断是否满足复用原相机设置的条件。
• 支持使用原始生成 PLY 初始化，旧图与新图共同使用正常完整训练配方，不再采用补拍专用的缩短步数或学习率倍率。
• 视频补拍沿用正常抽帧逻辑，所有结果写入独立派生项目，不覆盖源项目、原模型或编辑结果。
• 补齐准备阶段日志和最终发布数量，修复完成后仍显示“等待结果校验”的问题。

原模型初始化不是完整优化器断点续训，也不保证只改变补拍区域。普通与全景投影仍不能混合补拍。

图片 GPS 辅助生成

• 自动识别 JPG、JPEG、PNG 内嵌 EXIF／XMP GPS，并在素材区域提示。
• 有效定位信息可辅助匹配和相机重建；条件可靠时提供估计北向和米制尺度。
• 信息不足或定位结果校验未通过时，继续使用不依赖 GPS 的原有流程。
• GPS 原始坐标保存在本地，不加入统计或诊断上传。

当前仅在已验证的 Windows x64 路径启用。GPS 不代替视觉匹配，不承诺测绘精度；真实全景、跨设备补拍及其他平台仍有待验收。

预览性能与漫游

• 画面稳定后停止无意义重绘，交互与动画预览最高60 FPS，保持原有分辨率和模型细节。
• 主动旋转视角时平滑消除相机侧倾，不旋转网格或模型；自动定位仍保留真实训练相机朝向。
• 右下角新增0.1×～5.0×漫游速度滑轨，调整立即生效，并保存本机偏好。
• 编辑、撤销／重做及主动提交数值后，可直接继续 WASD 水平漫游，无需再次点击画布。
• 保留随缩放调速、Shift 三倍加速及输入框、弹窗等操作隔离。

WASD 仍仅在透视调整模式可用；正交和三视图不会响应。

导出与缓存限制调整

• 取消压缩 PLY、SPZ、SOG 的估算内存预检门槛。
• 取消全景视频和图片缓存固定20 GiB限制，保留单图解码、尺寸、溢出及缓存完整性检查。
• SPZ 坐标越界时暂停导出，用户可取消，或确认仅对导出副本居中、必要时等比缩小。
• 自动调整不修改原项目；导出结果会说明采用的居中参考点和缩放比例。
• 保留取消、源文件一致性校验及安全发布，失败时保护已有导出。

取消人为门槛不代表资源无限。SPZ 自动缩放可能改变尺寸、单位和精度；无法安全编码时仍会停止导出。

引擎检测、错误诊断与统计

• 改进 FFprobe 预检及引擎健康详情，右上角使用简短状态，悬停可查看具体检测结果。
• 区分训练显存不足、GPU资源限制、训练数据读取失败及 COLMAP 异常终止，避免后续错误覆盖原始原因。
• 保留实际退出信息，不再将进程崩溃默认归因为素材无法重建。
• 补齐补拍、失败前已知数据、最后执行阶段及初始／最终训练配置统计。
• 按次保存 Brush 实际训练设备信息，诊断报告继续沿用用户确认和原有上传协议。

本次修复错误识别与提示，不代表已经解决实际显存不足或引擎崩溃的底层原因。

💻 平台支持

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

COLMAP 可按平台和设备条件使用 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相关说明为准。

📦 安装包

• Windows：OOOSplat-0.7.1-x64-windows.exe
• Apple Silicon macOS：OOOSplat-0.7.1-arm64-macos.dmg
• Ubuntu 24.04：OOOSplat-0.7.1-x64-linux.deb

---

English

OOOSplat 0.7.1 updates generation defaults and Detailed settings, upgrades reshoots, adds image GPS assistance, frame-sampling controls and movement-speed adjustment, and improves bridge backfill, preview performance, exports, and diagnostics.

✨ Highlights

Generation Defaults and Automatic Retry

• Previous Auto Optimize settings are now standard, including automatic frame backfill and loop detection.
• All presets enable two-view triangulation only when the images contain actual transparency. Fully opaque Alpha images do not enable it.
• Auto Optimize is replaced by Automatic retry, enabled by default. It permits eligible retries with a lower splat ceiling or a different reshoot plan.
• After explicit GPU OOM, any preset may retry once at 80% of the failed process's last trustworthy splat count, without lowering training resolution.
• Completed details show “Automatically retried” only when a controlled retry actually ran.

Missing counts, ineffective reductions, or limits below initialization fail directly. Initialization points are not silently discarded. Disabling Automatic retry does not disable normal backfill or backend compatibility fallbacks.

Frame Sampling and Bridge Backfill

• Adds a stepped sampling slider with manual adjustment and Reset to default. Fast / Balanced / Detailed retain initial targets of 6 / 8 / 12 FPS.
• New runs with a custom ratio may add at most 20% of the actual initial frame count, capped by available source frames. For example, 360 initial frames permit at most 72 extra.
• Initial registration below 80% with a valid initial model can trigger one targeted bridge attempt in unregistered intervals, without duplicating existing frames.
• Higher manual ratios can still use remaining source frames; 100% sampling adds none. Initial frame indices and the 30-frame minimum remain unchanged, with all frames used when fewer are available.
• Default sampling retains the 9 / 12 / 15 FPS candidate budgets for Fast / Balanced / Detailed.
• Panorama bridges reuse real timestamps already read during generation, including VFR and nonzero start times, without probing the entire video again.
• Candidate selection avoids building a full source-frame array solely for backfill. Adjusting the ratio replans without repeating media detection.

The 1.2× ceiling limits frame counts, not runtime or success probability. Interrupted older runs retain their saved rules; manual retries of failed tasks use the new policy.

Detailed Training Settings

• Detailed training resolution no longer uses VRAM tiers. Ordinary inputs train at source resolution.
• Panorama Detailed retains native ERP and derives training faces at floor(ERP width / 4), without the previous 1024-pixel face cap or upscaling.
• Ordinary Detailed budgets splats from effective pixels and mean reconstruction track length, with coefficient 0.5. Fully transparent pixels are excluded.
• Panoramas retain separate point budgets. Actual settings follow the saved task configuration and are not multiplied by six.

Fast and Balanced retain their existing resolution policies. Ceilings do not guarantee achieved counts or quality. Native-resolution training may substantially increase VRAM, RAM, and disk requirements.

Reshoot Workflow

• Ordinary reshoots support different resolutions, aspect ratios, devices, and lenses. Panorama projects accept full ERP video and image reshoots.
• Simple questions select the processing approach and determine whether the original camera settings can safely be reused.
• Supports initialization from the original generated PLY, training old and new views with the normal full recipe instead of shortened steps or reshoot-specific learning-rate multipliers.
• Video reshoots use normal frame sampling. Results go to a separate derived project without overwriting the source project, model, or edits.
• Adds preparation logs and verified published counts, and fixes the lingering “Waiting for result validation” state after completion.

PLY initialization is not a full optimizer resume and does not guarantee changes remain limited to new regions. Mixing ordinary and panorama projections remains unsupported.

Image GPS Assistance

• Detects embedded EXIF/XMP GPS in JPG, JPEG, and PNG images and shows a notice in the media area.
• Valid locations can assist matching and reconstruction. Reliable observations provide estimated north orientation and metric scale.
• Insufficient information or rejected alignment results retain the original visual-only workflow.
• Raw GPS coordinates stay local and are excluded from telemetry and diagnostic uploads.

Currently enabled only on the validated Windows x64 path. GPS does not replace visual matching or provide surveying guarantees. Real panorama, cross-device reshoot, and other-platform acceptance remains incomplete.

Preview Performance and Navigation

• Stable views stop unnecessary redraws. Interaction and animation run at up to 60 FPS while retaining resolution and model detail.
• Manual orbit smoothly levels camera roll without rotating the grid or model. Auto locate view retains the actual training-camera orientation.
• Adds a bottom-right movement-speed slider from 0.1× to 5.0×, applied immediately and saved as a device preference.
• WASD works directly after edits, undo/redo, and explicit numeric submission, without another canvas click.
• Retains zoom-relative speed, Shift ×3 acceleration, and input/dialog isolation.

WASD remains available only in perspective Adjust mode, not orthographic or fixed views.

Export and Cache Limits

• Removes estimated-memory gates for compressed PLY, SPZ, and SOG exports.
• Removes the fixed 20 GiB panorama video/image cache cap while retaining per-image decode, dimension, overflow, and cache-integrity checks.
• Out-of-range SPZ exports pause for confirmation. Cancel, or adjust only the export copy by centering and, when necessary, uniformly scaling it.
• The original project remains unchanged. Export results report the center reference and scale.
• Retains cancellation, source-consistency checks, safe publication, and protection for existing exports on failure.

Removing artificial gates does not remove actual resource limits. SPZ scaling may change dimensions, units, and precision. Export still stops when safe encoding is impossible.

Engine Health, Diagnostics, and Telemetry

• Improves FFprobe preflight and engine-health details, with a compact status and hover panel.
• Distinguishes GPU OOM, GPU resource limits, training-data read failures, and abnormal COLMAP termination without letting later errors overwrite the root cause.
• Preserves actual exit information instead of treating an unexplained crash as a source-reconstruction failure.
• Completes reshoot metrics, known pre-failure data, last-stage tracking, and initial/final training configuration.
• Saves actual Brush device information per attempt. Diagnostic uploads retain user confirmation and the existing protocol.

These changes improve classification and guidance; they do not fix the underlying causes of GPU OOM or engine crashes.

💻 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 as supported. Brush training still requires an available graphics backend. Consult the feature notes for platform-specific validation coverage.

📦 Packages

• Windows: OOOSplat-0.7.1-x64-windows.exe
• Apple Silicon macOS: OOOSplat-0.7.1-arm64-macos.dmg
• Ubuntu 24.04: OOOSplat-0.7.1-x64-linux.deb

---

This discussion was created from the release v0.7.1.