New Features | UniFab RTX Rapid Upscaler AI Features in Detail

Technical Details Analysis

• Network Structure: The main body of the model is a 40-layer deep residual convolutional network, embedded with 8 groups of Residual Blocks, each containing two layers of 3x3 convolution and ReLU activation. The number of channels is dynamically adjusted from 64 to 256 to meet different resolution requirements.
• Attention Module: Adopts a combined design of convolutional attention layer and SE (Squeeze-and-Excitation) module to enhance feature selection and context understanding capabilities. Spatial attention utilizes 3x3 convolution to capture neighborhood correlations and effectively extract local texture information.
• Denoising Module: Utilizes the Bidirectional Long Short-Term Memory Network (Bi-LSTM) structure to capture the temporal characteristics of noise distribution in the video, thereby achieving self-adaptation denoising and reducing reconstruction artifacts.
• Optimization Algorithm: During the training process, Perceptual Loss combined with L1 Loss is adopted to ensure that the restored image is not only accurate at the pixel level but also takes into account the naturalness and detail restoration perceived by the human eye.

Through multi-level technology integration, UniFab RTX Rapid Upscaler AI breaks through the bottlenecks of traditional super-resolution technology, achieving efficient and accurate detail restoration and image quality improvement, bringing users a clearer, more natural, and more immersive video experience. 

Artifact Removal Technology

During video compression, especially at low bitrates, color banding and blocking artifacts degrade image quality with unnatural colors and harsh transitions. These issues worsen during super-resolution and sharpening, further distorting the image. UniFab RTX Rapid Upscaler AI features a specialized artifact removal module that uses deep neural networks to detect and repair these compression artifacts, significantly improving image naturalness and smoothness.

Key Technology Innovation

Multi-scale Feature Fusion and Residual Learning

The RTX Rapid Upscaler AI artifact removal module uses Multi-scale Feature Fusion to capture artifacts of various sizes and shapes. Its CNN-based model extracts features at multiple spatial scales and employs Residual Connections to focus on residual artifacts, enhancing accuracy and generalization. This allows the model to detect both obvious block artifacts and subtle compression traces.

Edge-preserving Filtering Strategy

To prevent blurring and detail loss during artifact removal, the model uses an Edge-preserving Filtering mechanism. By combining Guided or Bilateral Filtering with a depth feature map, it selectively targets artifact areas while preserving edges and textures, ensuring delicate and natural restoration.

Sub-domain processing of color space

RTX Rapid Upscaler AI uses color space conversion to transform images from RGB to YCbCr, allowing separate processing of channels. It applies targeted artifact removal: restoring details and structure in the luminance (Y) channel, and addressing color banding and blocking in the chrominance (Cb, Cr) channels, greatly enhancing color transition quality.

Gradient Consistency Constraint

To maintain natural color gradients and reduce color banding, the model incorporates a Gradient Consistency Constraint during training. This regularizes the restored image’s gradient field, preventing stepped or striped color artifacts, ensuring smooth color transitions, and enhancing visual realism.

Temporal Filtering and Optical Flow-Assisted Smoothing

Artifact removal must ensure both single-frame quality and video temporal consistency. RTX Rapid Upscaler AI uses Temporal Filtering and Optical Flow Estimation to leverage motion between frames, enabling smooth artifact repair and reducing flicker and jitter. This ensures stable, coherent visuals and greatly improves dynamic scene viewing.

Self-Adaptation Suppression Mechanism

By adopting Self-Adaptation weight adjustment based on the attention mechanism, the model can intelligently discriminate the severity of artifacts in different regions, dynamically adjust the removal intensity, avoid over-smoothing in regions rich in details, and enhance the sharpness and realism of the image.

Technical Details Analysis

• Network Structure: The model is a deep residual CNN with a multi-scale encoder-decoder. The encoder extracts multi-level features, while the decoder uses upsampling and feature fusion to reconstruct artifact-free images.
• Loss Function: Training employs a composite loss combining pixel-level L1 loss, perceptual loss, and gradient consistency loss to enhance artifact removal and preserve natural details.
• Temporal Consistency Module: Uses optical flow to generate motion-compensated frames for temporal filtering and dynamic weight learning, enabling smooth cross-frame artifact removal and reducing video flicker.

By integrating multiple advanced technologies, UniFab RTX Rapid Upscaler AI effectively addresses amplified compression artifacts in traditional super-resolution, ensures true restoration and smooth detail transitions, and significantly enhances both visual experience and video content value.

Supports importing HDR and Dolby Vision video sources 

With the widespread adoption of HDR and Dolby Vision in film and TV, UniFab RTX Rapid Upscaler AI offers full support for these video formats, enabling seamless import and high-quality processing of diverse high dynamic range content—greatly enhancing visual experience and expanding application possibilities.

Technical Innovation Points

End-to-End High-Precision HDR Decoding and Color Management

UniFab accurately parses HDR10, HLG, and Dolby Vision metadata using advanced HDR decoding algorithms. It applies gamut mapping to convert Rec.2020 color space to display-specific gamuts like DCI-P3 and Rec.709, ensuring precise color reproduction and visual consistency across devices.

Dolby Vision Dynamic Metadata Processing

UniFab features a dedicated module for parsing Dolby Vision’s scene- and frame-based dynamic metadata. Combined with AI analysis, it automatically adjusts tone and luminance mapping to optimize exposure and color, delivering the best image quality on various displays.

Multi-Level Tone Mapping and Inverse Mapping

During import, UniFab uses multi-level tone mapping and inverse tone mapping to convert HDR videos to HDR or SDR curves accurately, balancing contrast and detail preservation. The inverse mapping also intelligently upgrades SDR content to HDR, complementing super-resolution for enhanced dynamic range.

High-Precision HDR10+ and Dolby Vision Metadata Synchronization

The system simultaneously extracts and processes HDR10+ and Dolby Vision metadata during import, integrating with real-time rendering pipelines to maintain metadata accuracy. This prevents color shifts and luminance drift, ensuring stable and visually accurate enhancement results.

Specific technical details

Compatibility of Video Containers and Encoding Formats

Supports HDR and Dolby Vision video content in encoding formats including HEVC (H.265), AV1, and VP9, and is compatible with a wide range of video container formats such as MP4, MKV, and MOV, ensuring flexible and diverse video input sources.

Dynamic Metadata Parsing Module

This module employs patent-grade real-time parsing technology, based on the dynamic metadata stream of inter-frame and scene changes, to calculate optimization parameters in real time and drive the AI enhancement algorithm to make detailed adjustments to brightness, contrast, and color.

Color Space Conversion and Management

Built-in high-precision color space conversion function library, supporting standards such as Rec.2020, Rec.709, DCI-P3, and BT.2100, combined with color Look up table (LUT) correction to ensure accurate mapping of color and brightness for each pixel.

HDR and Dolby Vision Metadata Synchronization Strategy

Through timestamp alignment and cache optimization, ensure precise synchronization between metadata and video frames, avoid image quality issues caused by transmission or processing delays, and smoothly present dynamic tone changes.

With deep support and innovative processing of HDR and Dolby Vision content, UniFab RTX Rapid Upscaler AI delivers robust technical capabilities for professional video production and premium viewing experiences, allowing users to fully unlock high dynamic range’s visual potential for richer, more vibrant image quality.

Performance and Effectiveness Evaluation

UniFab RTX Rapid Upscaler AI leverages a highly optimized neural network architecture alongside NVIDIA RTX GPU hardware acceleration to deliver efficient, stable video enhancement solutions. It boosts image quality while precisely managing latency and resource use, striking an ideal balance between performance and user experience across diverse hardware setups.

Tensor Core Mixed Precision Inference Optimization

RTX Rapid Upscaler AI fully utilizes NVIDIA Tensor Cores with FP16 and FP32 mixed precision training and inference, reducing memory bandwidth and power usage while maintaining accuracy. This boosts inference speed, enabling 4K video processing at over 160fps for real-time playback and interactive editing.

Integrated Model and Computational Graph Fusion

The model employs fusion compilation to combine multiple CNN operations (convolution, activation, normalization) into a single efficient kernel, minimizing memory access frequency and latency. Coupled with CUDA graphs and NVIDIA’s memory management, this optimizes GPU resource use, balances video memory and compute load, and ensures system stability in multitasking.

Dynamic Load Balancing and Resource Scheduling  

For multi-GPU and multitask scenarios, an intelligent scheduling algorithm dynamically allocates resources based on GPU utilization, memory bandwidth, and task priority, preventing bottlenecks and resource contention to ensure smooth video processing.

Multi-Frame Temporal Consistency Acceleration

Using optical flow estimation, the system jointly processes multiple frames with optimized temporal fusion via parallel pipelines, enabling feature sharing and synchronized processing. This reduces redundant computations, enhances temporal image stability, and minimizes artifacts and jitter for smoother visuals.

Specific performance

• Processing Speed: On an NVIDIA GeForce RTX 4060, RTX Rapid Upscaler AI delivers over 160fps for 4K super-resolution, vastly outperforming traditional CPU and non-accelerated GPU solutions.
• Image Quality Improvement: Subjective and objective assessments show significant gains in PSNR and SSIM compared to traditional methods, outperforming bilinear interpolation and conventional super-resolution. Color reproduction is more natural, and artifact suppression is greatly enhanced with better detail restoration.
• Temporal Stability: The multi-frame optical flow and artifact removal mechanism reduces flicker and jitter. Visual Multi-Method Assessment Fusion (VMAF) shows over 15% improvement in temporal stability, ensuring smooth and consistent playback.

Future Outlook