With the release of **DepthAI 3.4.0**, we’re introducing an exciting new capability for OAK4 devices: **High Frame Rate (HFR) mode**. This feature pushes the limits of real-time perception, enabling **up to 480 frames per second**—while still running neural networks at that same blazing speed. This is not just high-speed capture—it’s **high-speed AI**. [upl-image-preview uuid=e0aa04f3-9402-4a43-be23-ede87cd86ef8 url=https://discuss.luxonis.com/assets/files/2026-04-10/1775806535-644003-hfr-optimized.gif alt={TEXT?}] ## **What’s New: High Frame Rate Mode** HFR mode is currently available on **OAK4 devices powered by the RVC4 platform**, using the **IMX586 sensor**. It represents an early preview of what’s possible with next-generation DepthAI pipelines. With HFR enabled, you can: * Capture ultra-fast motion with minimal blur * Run neural networks on every frame * Achieve ultra-low latency perception loops ## **Supported HFR Resolutions** At launch, HFR mode supports two fixed configurations: * **1920 × 1080 @ 240 FPS** * **1280 × 720 @ 480 FPS** These modes are currently fixed—resolution scaling and arbitrary FPS selection are not yet supported. However, additional flexibility is planned in future releases. In the meantime, you can use **ImageManip** to adapt outputs as needed. ## **Neural Networks at 480 FPS** One of the standout features of HFR mode is the ability to run neural networks at full frame rate. For example, the object detection demo runs **YOLOv6 at 480 FPS**, enabling real-time detection in extremely fast-moving environments. This opens up new possibilities for: * Industrial automation * Robotics requiring rapid response * Sports analytics * High-speed inspection systems ## **Example Applications** The release includes three example pipelines demonstrating HFR capabilities: ### **1\. Object Detection (YOLOv6 at 480 FPS)** Runs real-time object detection using HFR input. [https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr\_nn.py](https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_nn.py) --- ### **2\. Small Live Preview** Displays a lightweight preview stream at 240 or 480 FPS. [https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr\_small\_preview.py](https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_small_preview.py) --- ### **3\. Video Encoding** Encodes and saves high frame rate video streams. [https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr\_save\_encoded.py](https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_save_encoded.py) --- ## **Performance Insights** All examples include a **BenchmarkIn node**, which reports runtime performance and latency. Here’s a sample output: \[2025-08-14 23:31:49.487\] \[ThreadedNode\] \[warning\] FPS: 474.3766 \[2025-08-14 23:31:49.487\] \[ThreadedNode\] \[warning\] Messages took 1.0118543 s \[2025-08-14 23:31:49.487\] \[ThreadedNode\] \[warning\] Average latency: 0.05904912 s This demonstrates: * **Near-maximum throughput (\~480 FPS)** * **Low end-to-end latency (\~59 ms)** * Stable, real-time pipeline execution --- High frame rate perception is no longer just about capturing more frames—it’s about **understanding more, faster**.

High Frame Rate Neural Networks on OAK4 at 480 FPS

Janez

With the release of DepthAI 3.4.0, we’re introducing an exciting new capability for OAK4 devices: High Frame Rate (HFR) mode. This feature pushes the limits of real-time perception, enabling up to 480 frames per second—while still running neural networks at that same blazing speed.

This is not just high-speed capture—it’s high-speed AI.

What’s New: High Frame Rate Mode

HFR mode is currently available on OAK4 devices powered by the RVC4 platform, using the IMX586 sensor. It represents an early preview of what’s possible with next-generation DepthAI pipelines.

With HFR enabled, you can:

Capture ultra-fast motion with minimal blur
Run neural networks on every frame
Achieve ultra-low latency perception loops

Supported HFR Resolutions

At launch, HFR mode supports two fixed configurations:

1920 × 1080 @ 240 FPS
1280 × 720 @ 480 FPS

These modes are currently fixed—resolution scaling and arbitrary FPS selection are not yet supported. However, additional flexibility is planned in future releases. In the meantime, you can use ImageManip to adapt outputs as needed.

Neural Networks at 480 FPS

One of the standout features of HFR mode is the ability to run neural networks at full frame rate.

For example, the object detection demo runs YOLOv6 at 480 FPS, enabling real-time detection in extremely fast-moving environments. This opens up new possibilities for:

Industrial automation
Robotics requiring rapid response
Sports analytics
High-speed inspection systems

Example Applications

The release includes three example pipelines demonstrating HFR capabilities:

1. Object Detection (YOLOv6 at 480 FPS)

Runs real-time object detection using HFR input.

https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_nn.py

2. Small Live Preview

Displays a lightweight preview stream at 240 or 480 FPS.

https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_small_preview.py

3. Video Encoding

Encodes and saves high frame rate video streams.

https://github.com/luxonis/depthai-core/blob/v3.4.0/examples/python/HFR/hfr_save_encoded.py

Performance Insights

All examples include a BenchmarkIn node, which reports runtime performance and latency. Here’s a sample output:

[2025-08-14 23:31:49.487] [ThreadedNode] [warning] FPS: 474.3766
[2025-08-14 23:31:49.487] [ThreadedNode] [warning] Messages took 1.0118543 s
[2025-08-14 23:31:49.487] [ThreadedNode] [warning] Average latency: 0.05904912 s

This demonstrates:

Near-maximum throughput (~480 FPS)
Low end-to-end latency (~59 ms)
Stable, real-time pipeline execution

High frame rate perception is no longer just about capturing more frames—it’s about understanding more, faster.