This blog post was originally published at Arm’s website. It is reprinted here with the permission of Arm.
Discover how SME2 brings flexible, high-performance AI denoising to mobile photography for sharper, cleaner low-light images.
Every smartphone photographer has seen it. Images that look sharp in daylight but fall apart in dim lighting. This happens because signal-to-noise ratio (SNR) drops dramatically when sensors capture fewer photons. At 1000 lux, the signal dominates and images look clean. At 1 lux, readout noise appears as grain, color speckles, and loss of fine detail.
That is why neural camera denoising is one of the most critical and computationally demanding steps in the camera pipeline. When done well, it transforms noisy frames into sharp, vibrant captures. When done poorly, it leaves smudges and artifacts that ruin the shot.
Arm Scalable Matrix Extension 2 (SME2) advances denoising on mobile. It is a powerful new technology for CPU-based AI inference that is enabled across our new C1 CPUs. It is featured in several new flagship smartphones, see the device list.
SME2 is designed to accelerate a range of AI operations from generative AI to computer vision. It enhances the latest computational photography experiences. This brings automated image improvements for sharper, cleaner images with unprecedented speed and efficiency.
In this blog post, we explain how this happens.
Scalable Matrix Extensions for imaging innovation
Dedicated image signal processor (ISP) hardware remains highly effective for standard tasks such as denoising, demosaic, and tone mapping. However, imaging algorithms are evolving rapidly, and fixed-function blocks cannot easily adapt.
Arm Scalable Matrix Extension 2 (SME2) adds a new layer of flexibility. SME2 combines wide SIMD and matrix-multiply compute capability, enabled by Arm’s SVE2 (Scalable Vector Extension 2) and SME ISA features.
This combination enables high-throughput AI and computer vision (CV) acceleration directly into the CPU pipeline. Making it easier to integrate new algorithms without waiting for hardware refreshes.
SME2-enabled C1 CPUs enable OEMs and developers to:
- Match or exceed DSP-level performance in imaging workloads.
- Run some applications without using separate AI accelerators, thanks to SME2’s scalable throughput.
- Benefit from a CPU-like programming model, making it easier for developers to optimize and evolve code.
Neural camera denoising on SME2-enabled C1 CPUs
Arm has developed a neural camera denoising pipeline purpose-built for SME2. It operates directly in the RAW domain for superior noise modeling and detail retention.
It is built from two complementary algorithms:
UltraLite
- Temporal
- Downscale, per-channel processing, motion mask estimation, temporal accumulation.
- Efficient; stabilizes video in low light.
CollapseNet
- Spatial
- Cascaded, pyramid-based denoiser (UGGV color space)
- Superior detail retention in sub-lux conditions
When combined, UltraLite and CollapseNet form a spatio-temporal denoising pipeline. UltraLite delivers both temporal stability and CollapseNet restores spatial detail.
This combination ensures versatility. UltraLite excels at video, while CollapseNet ensures high-quality stills. Together, they provide robust denoising across the full range of scenarios.
Real-time performance on a single core
Neural camera denoising achieves real-time throughput, even when running on a single CPU core with SME2 enabled. The table below shows how SME2-enabled CPUs balance efficiency and flexibility to deliver DSP-class performance without requiring separate accelerators.
| UltraLite (temporal only) | 1080p | >180fps | Lightweight, efficient temporal denoising |
| CollapseNet (spatial) | 4K | ~30fps | High-quality RAW-domain denoising |
| Combined (spatio-temporal variant) | 4k | ~30fps | UltraLite + CollapseNet pipeline for video and stills |
Programmability and developer tools
Neural camera denoising is implemented as optimized C++ code and includes standalone benchmarking binaries for aarch64 targets. Developers can provide custom inputs, measure performance, and debug with ease.
Crucially, SME2 supports Arm C Language Extensions (ACLE) intrinsics. This allows
- Low-level tuning of critical kernels, such as convolutions and blending.
- Familiar workflows, using the same toolchains developers already rely on for Arm CPUs.
For experimentation, PyTorch and Keras models are also available. This enables rapid prototyping before deploying optimized implementations.
Explore the open-source code in the KleidiAI Camera Pipelines repository on GitLab.
Results: Extending image quality
Lab evaluations show that SME2-based neural camera denoising improves image quality in the conditions that matter most; 1 lux and below. In these low-light conditions, SME2-based denoising produces sharper, cleaner, and more natural results than ISP-only pipelines or even premium handsets
This highlights SME2’s complementary role. It works alongside the ISP, and takes over when fixed-function hardware reaches its limits.
Looking ahead
Neural camera denoising is only the beginning. SME2 also accelerates cinematic mode (depth-of-field effects), low-light enhancement, and other advanced camera features. The combination of performance, programmability, and scalability positions SME2 as a general-purpose imaging accelerator. It complements ISPs and enables continuous software innovation.
Conclusion
Noise is one of the hardest problems in photography. Low-light conditions push sensors to their limits. SME2-enabled C1 CPU neural camera denoising gives device makers a flexible, high-performance tool to deliver superior low-light imaging. It acts not as a replacement for ISP hardware, but as a complementary capability that extends what cameras can do.
SME2 combines ACLE programmability, real-time 4K performance on a single core, and open-source examples available today. Together, these make it a powerful technology for the next generation of computational photography.
Importantly, SME2 demonstrates the power of hardware and software algorithm co-design, where silicon capabilities and software techniques evolve together to unlock entirely new imaging possibilities.
Try it today with the AI Camera Pipelines on GitLab:
AI Camera Pipelines repository on GitLab
David Packwood
Principal Computer Vision Architect, Arm
