- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces UENR-600K, a synthetic paired dataset for nighttime video deraining. The authors model rain as 3D particles inside Unreal Engine 5 scenes, producing nighttime rain effects with chromaticity, localized illumination, glimmer, rain curtains, and scene occlusion. The dataset contains 600,000 paired 1080p rainy and clean frames. The authors also adapt Wan 2.2-TI2V-5B to video-to-video deraining using condition tokens, generation tokens, shared positional embeddings, an attention mask, and flow matching. Six existing methods and the Wan baseline are evaluated using PSNR/SSIM on synthetic data, VLM and human preferences on real nighttime rain videos, and temporal-consistency metrics. The paper claims that training on UENR improves real-video deraining and that the Wan baseline performs best under its evaluation protocol.
Critical strengths
The dataset addresses a specific limitation of nighttime deraining data: rain visibility and appearance depend on nearby lights and scene geometry. Its paired videos capture effects that global rain overlays represent poorly. This makes the dataset a useful resource for studying nighttime restoration.
The experiments test the data across several model families, rather than relying solely on the proposed Wan baseline. Synthetic metrics, real-video judgments, and the additional WeatherWeaver comparison provide evidence of its practical utility. The Wan adaptation also supplies a strong baseline, with training details and inference costs reported.
Critical weaknesses
The experiments do not fully isolate the source of the gains. UENR has about 42 times as many training frames as SynNightRain. Appendix F addresses scale through an equal-frame WeatherWeaver comparison, but illumination and restoration targets also differ. Appendix A does not establish scene or trajectory independence between training and testing. Likewise, Wan’s pretraining, full-resolution input, and 90-frame context complicate attributing its advantage to the architectural adaptation. These limitations weaken the causal explanations more than the dataset’s demonstrated utility.
The real-video evaluation needs clearer documentation. Annotator and sample counts, presentation format, and aggregation are missing from the human-study description. Table 3 reports 95.2% human preference and a 95.5% evaluator average, while the supplementary project page reports 97.6% and 96.3%. The temporal VLM examines only four midpoint frames, so its score supports local consistency rather than consistency throughout the clip. These issues limit confidence in the precise preference rate and the breadth of the temporal claims.
Questions for authors
How many annotators and evaluation items were used, what did annotators view, and how were judgments randomized and aggregated? Please reconcile the PDF and project-page preference rates and provide the underlying counts.
Are the training and test frames independent in scene and camera trajectory? Why was Kyoto excluded? Beyond the equal-frame WeatherWeaver comparison, what evidence separates the effects of rain rendering from scale, illumination, and target differences?
Which Qwen model and evaluation settings were used? How representative are the four midpoint frames of temporal behavior across the full clip?
How much of Wan’s advantage can be attributed to the adaptation beyond pretraining and larger spatial and temporal context? Why were NightRain (Lin et al., AAAI 2024) and ASF-Net (Xue et al., Pattern Recognition, 2025) excluded from the comparisons?
6 comments
Q1.1: Please provide selection counts for each evaluator and method in Table 3, win/tie/loss counts for Table 2, and the denominator and tie-handling rules.
Appendix H: Raw Results provides the counts and individual human responses. Qwen multi-way counts are reconstructed from the archived table and its 124-comparison protocol.
- Table 2 denominator: wins + losses + ties.
- Table 3 denominator: the sum of selections across methods, with no tie category.
- Ties: converted directly to percentages, like wins and losses, without further processing.
Q1.2: Why do the PDF values of 95.2%/95.5% differ from the project-page values of 97.6%/96.3%?
95.2% in Table 3 is the final human-evaluation result. Thank you for pointing this out. We have revised the paper to quote this human result explicitly and will revise the project page as well.
Q2.1: Please describe the complete human-evaluation protocol.
- Number of annotators: 21.
- Videos or frames: videos.
- Judgments per annotator: 8 dataset comparisons (two-way, including ties) and 2 method comparisons (multi-way), giving 168 + 42 = 210 selections overall.
- Randomization: candidate order was randomized.
- Blinding: method names were hidden and candidates were shown anonymously. Figures 9 and 10 illustrate the two video-based evaluation tasks.
- Aggregation: selections were summed and converted to percentages, without further processing.
- Inter-rater agreement: To maximize randomization, evaluators judged different video comparisons rather than repeated, identical comparisons. We therefore cannot estimate inter-rater agreement. Individual responses are now provided in Tables 18 and 19 in Appendix H for direct inspection.
Q2.2: Why does Appendix A describe the real set as VLM-only while Tables 2/3 report human results?
The VLM-only description in Appendix A was a reporting error. Tables 2 and 3 are correct, and Appendix A now states that the real set is used for both VLM and human evaluation. Thank you for pointing this out.
Q2.3: Please specify the Qwen checkpoint, decoding settings, prompt and repeated runs, and justify using four midpoint frames for temporal evaluation.
- Qwen version and checkpoint:
Qwen3-VL-32B-Instruct. - Decoding settings: temperature 0.7, top_p 0.8, top_k 20 and repetition_penalty 1.0.
- Prompts: the same as Claude's, provided in Appendices C–D.
- Repeated runs: one run per item.
- Four midpoint frames: sparse sampling, such as one frame per second, leaves gaps too large to assess local flicker. We instead use four consecutive midpoint frames and three difference maps to evaluate sampled-window temporal consistency. Full-clip AFD complements this evaluation across all 89 adjacent frame pairs.
Q3.1: Was a matched-scale or matched-scene comparison performed?
Regarding dataset scale, we added a matched-frames experiment, training Our Baseline on 11,880 UENR frames, close to SynNightRain's 11,915. The remaining training settings match those in the paper.
We evaluated 372 frame pairs from 124 Pexels clips with Qwen3-VL-32B-Instruct (Appendix I.6). The UENR-trained model is preferred in 301 comparisons (80.9%), the SynNightRain-trained model in 24 (6.5%), with 47 ties (12.6%).
Q3.2: Are the 1,800 frames independent of training at the scene, virtual-environment and camera-trajectory levels?
As explained in W1, these are in-distribution validation frames, not the test frames used in the main real-video experiments. They share the training environment and continuous camera trajectory and measure reconstruction within that environment. The main experiments use the separate 11,160-frame Pexels real-video test set, with independent scenes and trajectories.
We have revised "1,800 test frames" to "1,800 validation frames" in the paper to avoid confusion.
Q3.3: Please explain why Kyoto is excluded from the benchmark.
We did not exclude Kyoto from the dataset. Its 100K paired frames are provided as additional resources for future research.
On two 90-frame Kyoto clips, the CitySample-trained baseline removes visible rain without further finetuning (Figure 19; Appendix I.4). Mean RGB PSNR is 22.57 dB, compared with 23.45 dB for the rainy inputs. Texture and brightness differences remain, so pixel-level fidelity does not improve on these clips.
Thank you for the constructive review. We address each concern below.
Regarding dataset scale, we added a matched-frames experiment, training Our Baseline on 11,880 UENR frames, close to SynNightRain's 11,915. The remaining training settings match those in the paper.
We evaluated 372 frame pairs from 124 Pexels clips with Qwen3-VL-32B-Instruct (Appendix I.6). The UENR-trained model is preferred in 301 comparisons (80.9%), the SynNightRain-trained model in 24 (6.5%), with 47 ties (12.6%).
We address the count and reporting questions in Q1.1–Q1.2 below, and the protocol and temporal-evaluation questions in Q2.1–Q2.3.
Role of pretraining:
We trained Our Baseline without its pretrained denoiser weights, using the same LoRA adaptation procedure. The VAE and text encoder remain pretrained. Both the SynNightRain and UENR runs produce noise-like outputs without recognizable scene structure on the evaluated Pexels clip (Figure 20; Appendix I.5). These short adaptation runs do not isolate pretraining under matched training budgets.
Isolating UENR's contribution: This is already provided by the dataset comparison in Table 2, specifically the "Our Baseline" row, rather than the method comparison in Table 3.
Regarding the positioning, UENR's novelty is the dataset as a whole, including UE5 in-scene lighting and occlusion, scale, and its benchmark. This is why the paper is titled "UENR-600K". Your summary is accurate and matches the contribution claimed in our paper.
Regarding the related works you mentioned: