跳转到内容

重建一致性

reconstruction_consistency 是公开指标 geometric_consistency 与 photometric_consistency 共用的计算单元。Depth-Anything-3(DA3)从预测帧恢复深度与相机位姿,构建彩色点云,并评估这些点跨视角重投影的一致性。它只依赖 prediction 本身,不需要 reference video 或几何真值。

与旧版 World-R1 的 Gaussian splat + LPIPS S_recon 不同,也与 WorldLens 中对预渲染 reconstruction 视频做光流 warp + LPIPS 的 novel_view_consistency 不同,本指标直接在预测帧上运行 DA3(infer_gs = false),由深度图、估计位姿与 RGB 帧构建点云,再检验跨帧重投影是否与预测深度一致。因此它衡量的是重建路径下的几何自洽性,而不是 prompt following、感知质量或 reference-video fidelity。

对每一帧,backend 将其他帧观测到的点重投影到当前视角,并计算几何误差(重投影深度与预测深度图的 median relative depth error)与光度误差(z-buffer 渲染颜色相对原始帧的 PSNR)。headline 分数为几何一致性:

geometric_consistency=11+mean(valid_geo_errors)(1)\mathrm{geometric\_consistency} = \frac{1}{1 + \mathrm{mean}(\mathrm{valid\_geo\_errors})} \tag{1}

光度 PSNR 采用标准映射:

photometric_consistency=1−10−PSNR/20.(2)\mathrm{photometric\_consistency} = 1 - 10^{-\mathrm{PSNR}/20}. \tag{2}

两个公开指标都落在 [0,1][0, 1],越高越好,并分别聚合。

案例按 Geometric Consistency 与 Photometric Consistency 各选三组;高分侧均有明显运动。每组分数只对应标注的指标。 Audit manifest: docs/assets/metric_examples/missing_pages_examples_20260824.json.

输入图像: image_static_stylized_verdant_landscape_0122
输入图像
输入已核验 · 同一 benchmark case · gap 0.8063

image_static_stylized_verdant_landscape_0122

套件: image_static指标: geometric_consistency场景: stylized / outdoor / verdant_landscape
基准 prompt target

Camera pan left. A verdant landscape with trees, extended terrace, and foliage.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 geometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.9909

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.1847
输入图像: image_static_stylized_suburb_0009
输入图像
输入已核验 · 同一 benchmark case · gap 0.7897

image_static_stylized_suburb_0009

套件: image_static指标: geometric_consistency场景: stylized / outdoor / suburb
基准 prompt target

Camera orbit right. A traditional courtyard with stone bench, garden elements, and traditional buildings.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 geometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.9848

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.1952
输入图像: image_static_photorealistic_city_0047
输入图像
输入已核验 · 同一 benchmark case · gap 0.8076

image_static_photorealistic_city_0047

套件: image_static指标: geometric_consistency场景: photorealistic / outdoor / city
基准 prompt target

Camera orbit left. A city street with tram, bicycle, and historic building.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 geometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.9894

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.1818
输入图像: image_static_stylized_city_0123
输入图像
输入已核验 · 同一 benchmark case · gap 0.3239

image_static_stylized_city_0123

套件: image_static指标: photometric_consistency场景: stylized / outdoor / city
基准 prompt target

Camera pull out. A park path with benches, trees, and wider path. Camera move right. A city street with buildings, parked cars, and benches. Camera pan right. An urban square with fountain, people, and shops.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 photometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.8311

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.5073
输入图像: image_static_photorealistic_passageways_0106
输入图像
输入已核验 · 同一 benchmark case · gap 0.2983

image_static_photorealistic_passageways_0106

套件: image_static指标: photometric_consistency场景: photorealistic / indoor / passageways
基准 prompt target

Camera pan right. A modern hotel corridor with elevator doors, emergency exit sign, and wall panels. Camera move right. A lobby area with reception desk, seating area, and decorative plants. Camera pull out. Entrance doors with revolving door and floor mat.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 photometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.9052

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.6070
输入图像: image_static_stylized_passageways_0090
输入图像
输入已核验 · 同一 benchmark case · gap 0.3448

image_static_stylized_passageways_0090

套件: image_static指标: photometric_consistency场景: stylized / indoor / passageways
基准 prompt target

Camera pan right. A stylized passageway with columns, wall sconces, and corridor.

prompt source: report_details

高分侧有明显的镜头或场景运动;两侧首帧均与同一输入图完成几何匹配。此处只比较 photometric_consistency。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.8870

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.5423