HPSv3-Norm

HPSv3-Norm measures human-preference-aligned visual quality of a generated video using the HPSv3 reward model. It is prediction-only: each evaluated frame is scored independently, without pairing against a reference rollout.

hpsv3_norm is an official image_static quality metric and enters both per_metric_summary and the equal-weight seven-metric quality total.

WorldAtlas uniformly subsamples up to 20 frames from the prediction video. For frame $x_i$, the HPSv3 inferencer outputs a raw reward $r_i$. The per-video mean raw reward is

$$\bar{r} = \frac{1}{N}\sum_{i=1}^{N} r_i \tag{1}$$

Let $p_1$ and $p_{99}$ denote fixed normalization percentiles (defaults: $p_1 = 5.21$, $p_{99} = 8.66$). The normalized score on $[0, 100]$ is

$$S_{\mathrm{HPS}} = \mathrm{clip}\!\left( \frac{\bar{r} - p_1}{p_{99} - p_1} \times 100,\, 0,\, 100 \right) \tag{2}$$

WorldAtlas stores the metric raw value in $[0, 1]$ as $S_{\mathrm{HPS}} / 100$. Higher values indicate stronger human-preference alignment. When the HPSv3 backend or checkpoints are unavailable, the metric is reported as not_applicable.

Why 1.0000 / 0.0000?

These are real evaluation results, not hand-entered placeholders. The gallery intentionally selects verified same-input pairs with the largest normalized score gaps. HPSv3-Norm first averages 20 per-frame raw rewards and then applies Equation (2): a mean at or above $p_{99}=8.66$ is clipped to 1.0000, while a mean at or below $p_1=5.21$ is clipped to 0.0000.

The three displayed pairs have the following raw reward means (higher / lower):

• Workspaces 0020: 8.872 / 4.902
• Verdant landscape 0125: 8.863 / 2.239
• Stylized living space 0142: 8.904 / 2.930

The raw rewards therefore differ across videos even though percentile normalization maps every displayed pair to the two clipped endpoints.

Verified Same-Input Examples

Audit manifest: docs/assets/metric_examples/curated_same_case_manifest.json.

Input image

Input verified · same benchmark case · gap 1.0000

image_static_photorealistic_workspaces_0020

Suite: image_staticScene: photorealistic / indoor / workspaces

Benchmark prompt target

Camera orbit left. A futuristic control room with swivel chairs, digital screens, and control panels.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 1.0000

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。

Lower metric scoreScore 0.0000

Input image

Input verified · same benchmark case · gap 1.0000

image_static_photorealistic_verdant_landscape_0125

Suite: image_staticScene: photorealistic / outdoor / verdant_landscape

Benchmark prompt target

Camera move left. Ancient ruins with side garden, stone bench, and wildflowers.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 1.0000

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。

Lower metric scoreScore 0.0000

Input image

Input verified · same benchmark case · gap 1.0000

image_static_stylized_living_spaces_0142

Suite: image_staticScene: stylized / indoor / living_spaces

Benchmark prompt target

Camera orbit left. A living room with sofa, coffee table, and bookshelves.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 1.0000

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。

Lower metric scoreScore 0.0000

References

HPSv3 — Human Preference Score

Vision-language reward model used to score image quality from human-preference supervision; WorldAtlas applies linear percentile normalization to per-video mean raw rewards.