How AV1 Frame Resizing Enables Dynamic Resolution
AV1 introduces native support for frame resizing directly within the encoded bitstream, allowing the video resolution to change dynamically on a frame-by-frame basis without inserting a keyframe. By enabling inter-frames to reference reconstructed frames of differing dimensions, the encoder can rapidly scale resolution up or down to manage bitrate fluctuations and network congestion smoothly. This article explains the technical mechanics of AV1's reference frame resampling, why it eliminates the need for expensive Intra (I-frame) resets, and how it optimizes modern real-time streaming workflows.
The Problem with Traditional Codecs
In legacy video formats such as H.264 (AVC) and standard implementations of H.265 (HEVC), the spatial resolution of a video sequence is strictly tied to the sequence parameter set. Changing the resolution requires the encoder to emit an Instantaneous Decoder Refresh (IDR) frame, also known as a keyframe.
Keyframes do not rely on previous frames for temporal prediction; they contain complete image data, making them significantly larger—often 5 to 10 times the size of a standard inter-frame (P-frame or B-frame). In scenarios with sudden network degradation, sending a massive keyframe to downscale the resolution worsens congestion, increases latency, and frequently leads to playback buffering or dropped frames.
Reference Frame Resampling in AV1
AV1 overcomes this limitation through a feature known as Reference Frame Resampling (or frame resizing). In AV1, individual inter-frames can be encoded at a different resolution than the frames stored in the reference buffer.
Instead of resetting the decoder buffer with a keyframe, AV1 allows an inter-frame to directly use a previously decoded frame of a different size as its motion-compensated predictor. The bitstream signals the new spatial dimensions at the frame header level, instructing the decoder to adapt without breaking the temporal prediction chain.
The Rescaling Process During Decoding
When an AV1 decoder encounters an inter-frame with dimensions that differ from its reference frame, it performs the following steps:
- Header Parsing: The decoder reads the new frame width and height signaled in the frame header.
- Internal Upscaling/Downscaling: Before calculating motion vectors, the decoder uses standardized, normative polyphase scaling filters to resample the stored reference picture to match the current frame’s spatial dimensions.
- Motion Compensation: Once the reference frame is scaled, the decoder proceeds with normal motion vector prediction and residual reconstruction.
Because the scaling algorithm is strictly standardized in the AV1 specification, both the encoder and decoder generate mathematically identical references, preventing visual drift between different decoder implementations.
Strategic Advantages for Real-Time and Adaptive Streaming
Dynamic frame resizing provides distinct operational advantages across various video delivery environments:
- Instantaneous Rate Control in WebRTC: In real-time communications, encoders can immediately downscale a frame to preserve target framerates during sudden bandwidth drops, then scale it back up when conditions improve—all without transmitting bandwidth-heavy keyframes.
- Seamless ABR Transitions: For video-on-demand and live streaming (DASH/HLS), switching between bitrate rungs traditionally causes visible transitions or requires tight keyframe alignment. Dynamic resolution changes allow a stream to transition smoothly across varying network conditions without waiting for fixed segment boundaries.
- Super-Resolution Modes: Encoders can intentionally downscale frames prior to encoding and signal the decoder to upscale them back to native resolution for final display. This preserves structural coherence in high-motion scenes where full-resolution encoding would otherwise introduce severe compression artifacts.