Abstract
Capturing high-quality video in extreme low-light conditions remains a significant challenge for mobile imaging systems due to the physical constraints of small sensors and short exposure times. The AIM 2025 Low-light RAW Video Denoising Challenge was established to address these limitations by providing a comprehensive benchmark featuring diverse sensor data and high-quality ground truth. This paper summarizes the challenge's findings, highlighting the critical role of temporal redundancy and metadata conditioning in bridging the performance gap in low-light environments.
Introduction
Low-light video capture on mobile devices is fundamentally limited by the relationship between frame rate and exposure time. Unlike static photography, where long-exposure techniques can mitigate noise, video is restricted to shutter speeds typically ranging from 1/24 to 1/120 of a second. In environments with illumination as low as 1 lux, sensors must employ high gain (ISO), which introduces significant read noise, photon shot noise, and various sensor-specific artifacts.
Existing denoising solutions often fail to generalize to these conditions because they are frequently trained on synthetic noise models or sRGB data. Such approaches ignore the linear, signal-dependent characteristics of RAW sensor data. The AIM 2025 Challenge addresses this by introducing a rigorous multi-device benchmark designed specifically for temporal RAW denoising, utilizing real-world data from a wide array of mobile sensors.
The AIM 2025 Dataset
A primary contribution of this challenge is the introduction of a massive, high-diversity dataset designed to simulate realistic mobile filming conditions.
- Scale and Diversity: The dataset comprises 756 ten-frame sequences captured across 14 different smartphone sensors, including Wide, Ultrawide, Telephoto, and Front-facing modules from Samsung, Google, and POCO devices.
- Capture Conditions: Data was collected under nine distinct combinations of illumination (1, 5, and 10 lux) and exposure settings (1/24, 1/60, and 1/120 s).
- Ground Truth Generation: To provide a "clean" target, high Signal-to-Noise Ratio (SNR) references were created by averaging 200–500 static burst frames. This process effectively eliminates stochastic noise, providing a gold standard that is otherwise impossible to capture in a single video frame.
- Precision Alignment: To solve the common issue of misalignment in video datasets, the authors utilized a motorized rail system to ensure repeatable and precisely aligned motion across sequences.
Methodology and Top-Performing Architectures
The challenge results demonstrated a clear trend: the most successful models were those that effectively exploited temporal aggregation rather than relying solely on spatial denoising.
1st Place: SNU-ISPL (DarkVRAI)
The winning team introduced the "DarkVRAI" framework, which centers on a Burst-Order Selective Scan (BOSS) module. This module is designed to capture long-range temporal dependencies across the video sequence. A key innovation in their approach was the use of "capture conditions" (ISO, illumination, and frame rate) as metadata. By using this metadata to condition the network via adaptive layer normalization, the model could dynamically adjust its denoising strategy based on the specific noise statistics of the capture environment.
2nd Place: XJAI
The runner-up team employed a multi-level hierarchical transformer. This architecture processed sequences of nine frames using a progressive refinement strategy. By leveraging the self-attention mechanism, the model was able to effectively manage inter-frame misalignment and varying noise levels across the temporal window.
Single-frame Baselines
Teams such as AxeraAI utilized high-performance convolutional models like NAFNet. While these models offer high efficiency and served as strong benchmarks, they ultimately highlighted the limitations of spatial-only processing in extreme low-light scenarios.
Key Insights and Results
The competition results provided several critical insights into the state of RAW video denoising:
- The Temporal Advantage: In the most challenging regimes (1 lux at 1/120s exposure), multi-frame methods outperformed single-frame approaches by more than 3 dB in PSNR. This confirms that temporal redundancy is essential for recovering signal in extreme noise.
- Contextual Awareness: The success of SNU-ISPL demonstrated that providing the model with explicit context (ISO and lux levels) significantly boosts performance, allowing for more precise adaptation to sensor-specific noise distributions.
- Real-world RAW Complexities: Participants noted that real-world RAW data often contains systemic biases, such as black-level errors and clipped negative noise values. These "dirty" characteristics can confuse standard Gaussian noise models, necessitating more robust architectural designs.
Quantitatively, SNU-ISPL achieved a top PSNR of 48.32 dB and an SSIM of 0.9879. XJAI followed closely with 48.19 dB. The gap between these temporal models and the best single-frame model (46.52 dB) was nearly 2 dB, a difference that was qualitatively visible in the recovery of fine textures and the absence of "oil painting" artifacts.
Limitations and Future Work
While the challenge represents a significant step forward, several hurdles remain for practical implementation:
- Calibration Robustness: Future models must better account for hardware-level calibration errors, such as black-level shifts that cause color casts in dark regions.
- Computational Efficiency: The reliance on complex transformers and temporal scanning poses a challenge for real-time mobile deployment (30+ fps). Optimization for Neural Processing Units (NPUs) is a necessary next step.
- Universal Generalization: Cross-sensor performance remains variable. Developing a universal denoiser capable of maintaining high performance on unseen sensor hardware remains an open research problem.
Summary updated