Existing web-generation benchmarks rely on text prompts or static screenshots as input. However, videos naturally convey richer signals such as interaction flow, transition timing, and motion continuity, which are essential for faithful webpage recreation. Despite this potential, video-conditioned webpage generation remains largely unexplored, with no dedicated benchmark for this task. To fill this gap, we introduce WebVR, a benchmark that evaluates whether MLLMs can faithfully recreate webpages from demonstration videos. WebVR contains 175 webpages across diverse categories, all constructed through a controlled synthesis pipeline rather than web crawling, ensuring varied and realistic demonstrations without overlap with existing online pages. We also design a fine-grained, human-aligned visual rubric that evaluates the generated webpages across multiple dimensions. Experiments on 19 models reveal substantial gaps in recreating fine-grained style and motion quality, while the rubric-based automatic evaluation achieves 96% agreement with human preferences. We release the dataset, evaluation toolkit, and baseline results to support future research on video-to-webpage generation.
Figures & tables
Figure 1: Overview of the WebVR benchmark. The left panel illustrates the task distribution across various webpage categories, while the right panel details the automated evaluation pipeline, which executes generated code in a standardized sandbox and scores it against human-aligned visual rubrics.
Figure 2: A case study illustrating the conversion from an original video into generated code and its rendered video output.
Figure 3: The WebVR data synthesis pipeline. The process consists of four stages: (A) Seed Data Preparation via semantic re-theming, (B) Visual Asset Retrieval to ground specifications, (C) Candidate Generation and Execution using multiple MLLMs, and (D) Automated Filtering and Refinement to construct the final high-quality benchmark set.
Figure 4: Statistics of the WebVR benchmark dataset, showing the distributions of (a) reference video durations (in seconds) and (b) the number of visual rubric items per instance.
Model
GA
NF
SSL
IM
Overall
Open-source Models
GLM-4.6V
22.78
15.42
7.17
14.35
11.42
Qwen3-VL-30B-A3B-Instruct
33.33
34.87
17.71
12.67
21.44
Qwen3-VL-30B-A3B-Thinking
53.38
60.47
33.49
20.22
37.69
Qwen3-VL-235B-A22B-Instruct
51.06
52.65
40.09
22.12
40.71
Qwen3-VL-235B-A22B-Thinking
61.20
68.04
43.11
29.30
46.80
Table 2: Results of the WebVR Benchmark. Joint evaluation based on code and rendered video, using Kimi-K2.5 as the evaluator. The best results are marked in orange , and the second best results are marked with an underline .
Figure 5: Hashed bars represent scores assigned without the visual rubric, while solid bars represent scores with the rubric applied. Note the inflated and compressed scoring distribution when the rubric is absent.
Evaluation Type
Gemini-3.0-Flash
Qwen3.5-397B-A17B
Kimi-K2.5
Average
Without Visual Rubric
Code Only
56.0
62.0
60.0
59.3
Video Only
62.0
66.0
72.0
66.7
Code and Video
58.0
58.0
76.0
64.0
Average
58.7
62.0
69.3
-
With Visual Rubric
Table 3: Agreement rates (%) between automated MLLM judges and human experts. Evaluated on 50 sampled instances, the table contrasts judge reliability with and without the visual rubric across different input modalities.
Figure 6: Representative failure cases across four models on the same webpage instance. Gemini-2.5-Flash produces a nearly invalid page with missing content. GLM-4.6V generates a structurally unrelated page that deviates entirely from the reference layout and theme. Qwen3-VL-235B-A22B-Thinking reproduces the overall structure but places images in incorrect locations with wrong references. Kimi-K2.5 achieves near-perfect reconstruction, with only minor inconsistencies in image overlap positioning.
Model
Temperature
Top-p
Top-k
Max_tokens
Extra penalties
GLM-4.6V
0.6
–
2
16k
repetition =1.1
Qwen3-VL-30B-A3B-Instruct
0.6
0.95
20
32k
presence =0.0 , repetition =1.0
Qwen3-VL-30B-A3B-Thinking
0.6
0.95
20
32k
presence =0.0 , repetition =1.0
Qwen3-VL-235B-A22B-Instruct
0.7
0.80
20
32k
presence =1.5 , repetition =1.0
Qwen3-VL-235B-A22B-Thinking
0.6
0.95
20
32k
presence =0.0 , repetition =1.0
Qwen3.5-VL-397B-A17B
0.6
0.95
20
32k
presence =0.0 , repetition =1.0
Table 4: Inference-time generation parameters for evaluated models on WebVR.