UniFab vs Topaz: A Comparison of Model Scale-Up Techniques

In video processing, AI-driven upscaling models are key to enhancing clarity and detail. UniFab and Topaz are leading software solutions excelling in video quality improvement, resolution upscaling, and noise reduction. This article compares their technologies, performance, effectiveness, and advantages across different applications.

Product Lines and Positioning

• UniFab: Offers a comprehensive platform covering video upscaling, noise reduction, HDR, and color enhancement, catering to a wide range of video enhancement and post-production needs.
• Topaz: Specializes in visual quality improvement, particularly video upscaling and detail restoration, powered by strong AI algorithms. Its flagship product, Video Enhance AI, is ideal for enhancing movies and short films.

Comparison of Topaz and UniFab Model Systems

Both UniFab and Topaz use advanced AI super-resolution algorithms, leveraging deep convolutional networks (CNN), attention mechanisms, and adversarial training (GAN) for video enhancement. While both excel at detail restoration, noise reduction, and sharpening, their model architectures differ significantly, leading to varied performance across different content types. To better understand these differences, we first compare their model architectures.

Topaz : A model system split based on "enhancement direction", Topaz's models are mainly classified according to functional focus, for example:

• Proteus: A general-purpose enhancement model with adjustable parameters, capable of fine-tuning noise reduction, sharpening, and various quality restoration parameters.
• Iris: Focuses on face enhancement, suitable for videos with high noise and face detail degradation caused by compression.
• Rhea: General enhancement, but biased towards detail restoration.
• Other models such as Theia / Nyx / Artemis: corresponding to different scenarios such as low noise, medium noise, and improvement of compression artifacts.

Features: Topaz's model emphasizes "enhancement methods" and "adjustable parameters", highlighting that users can adjust different loss weights to achieve different image quality optimization directions.

UniFab: A model system split based on "content type", UniFab provides more adaptable models according to the type of video creatives:

• Equinox: A general-purpose model suitable for daily creatives and mixed content (balancing speed and quality).
• Titanus (NEW / UniFab 4 Release): Designed specifically for film, TV series, and other film-grade creatives, with high dynamic range and optimized complex lighting effects.
• Kairo: Anime model, enhancing line, color block, and color consistency.
• Vellum: A texture enhancement model suitable for high-detail scenarios such as architecture, landscape, and creative photography.

Features: UniFab's model emphasizes "optimization for creative types", enhancing consistency and predictability, and reducing artifacts or frame breakdown issues caused by creative mismatches.

Overview and Technological Innovations of UniFab's Four Upscaler Models

Equinox — Balanced Enhanced Model

Positioning: Equinox is a balanced enhancement model designed for everyday video processing, prioritizing both speed and quality. It is especially suited for standard creative content, delivering high-quality enhancement with excellent real-time performance. Equinox excels in scenarios that demand fast feedback and efficient processing, making it versatile for various video enhancement needs.

Technical Features: Equinox utilizes Self-Adaptation resolution upscaling technology that dynamically adjusts processing based on content complexity. This approach maintains image quality while maximizing computing resource efficiency. Through optimized neural network architecture and inference strategies, Equinox reduces processing time and achieves well-balanced enhancement across diverse video types.

Core technological innovation of Equinox

Lightweight Neural Network Architecture Design

• Using variable-depth convolution, which decomposes standard convolution into depthwise convolution and pointwise convolution, significantly reduces the amount of computation and the number of parameters while maintaining efficient feature representation capabilities.
• Integrating the ShuffleNet module, it enhances feature mixing through the channel shuffle mechanism, improves the network's expressiveness and efficiency, and reduces memory access requirements.
• Designed a dynamic adjustment mechanism for network layer number and width, which adaptively adjusts the model scale according to the complexity of video content and optimizes the allocation of computing resources.

Multi-level Cache Hierarchy Optimization

• Design a multi-level cache system using on-chip cache to optimize data cache hit rate, reduce external video memory access, and lower latency and energy consumption.
• Combining near-storage computing technology, it moves part of the convolution computation closer to the storage unit, saving memory bandwidth and shortening the data transmission path.
• Implement a Self-Adaptation cache management strategy, dynamically schedule cache content, optimize cache allocation, and maximize cache utilization.

Prediction Pipeline Control

• Introduce a pipeline scheduling mechanism based on video frame content complexity prediction, evaluate computational load in advance, dynamically adjust pipeline depth and parallelism, and achieve on-demand allocation of computational resources.
• Predict and preprocess computational stalls and conflicts in the pipeline, reduce idle cycles through pipeline separation and rearrangement, and improve the utilization efficiency of computing units.
• Asynchronously overlap computation and data transfer to further reduce overall processing latency. 

Tensor Core Dynamic Compression Technology 

• Dynamically adjust the low-rank decomposition parameters of tensors in neural networks, automatically adjust the tensor rank according to the complexity of the current video frame, compress the tensor dimensions, and reduce the computational load.
• Combined with sparse tensor representation, redundant activations are removed through sparse activation pruning and channel sparsification strategies to improve inference speed.
• Implement dynamic conversion of tensor formats , leverage hardware-accelerated sparse matrix operations, and reduce memory access and computational burden. 

End-to-End System Collaborative Optimization

• Integrate lightweight network architecture, caching mechanism, pipeline control, and dynamic compression to build an end-to-end collaborative data processing mechanism, avoiding resource waste and performance bottlenecks.
• Optimize system performance, adapt to dynamic characteristics, and achieve optimal real-time energy efficiency ratio through computing unit scheduling and memory management at the hardware architecture level.
• Supports multi-threading, multi-core heterogeneous parallelism, and is compatible with GPUs and AI accelerators, ensuring that the model can achieve optimal performance on diverse hardware platforms. 

Application Scenarios and Performance

Equinox performs excellently in enhancing various standard video creatives, and is particularly suitable for scenarios that require rapid processing and feedback, such as:

• Short video editing
• Social Media Content Generation
• Enterprise Video Content Enhancement

With its lightweight network architecture and efficient computational resource management, Equinox can deliver outstanding performance in general video processing, helping users efficiently complete video enhancement tasks for general creatives. 

Vellum — Texture Enhancement Model

Positioning:Vellum is an efficient model focused on texture and detail enhancement, employing advanced technologies like spatio-temporal convolutional networks, self-attention mechanisms, and multi-scale feature fusion. It effectively restores complex textures and dynamic video changes while preserving naturalness and frame coherence. Vellum excels at upscaling and detail enhancement, ideal for high-texture scenarios such as architecture, landscapes, and fast motion.  

Technical Features and Innovations:The Vellum Texture Enhancement Model is built on a spatio-temporal convolutional network combined with self-attention and multi-scale feature fusion. It leverages intra-frame and inter-frame details through residual learning and multi-task loss optimization to accurately restore and efficiently enhance complex textures.

Core technological innovation of Vellum

Spatio-Temporal Convolutional Network 

• 3D Convolution Operation: By introducing 3D convolution kernels, Vellum can simultaneously extract features in both spatial and temporal dimensions, capturing dynamic changes and motion information in video sequences. The 3D convolution kernels slide between consecutive frames, fusing the texture and motion features of neighboring frames to effectively enhance the ability to represent motion details.
• Temporal Recursive Structure: By incorporating temporal recursive structures such as ConvLSTM or GRU, Vellum can dynamically adjust feature responses, highlight key motion regions and dynamic textures, and performs particularly well in scenarios with rapid motion and occluded changes. 

Self-attention mechanism

• Long-range dependency weighting: The self-attention mechanism effectively suppresses noise and enhances key structural information by calculating the interdependencies between different spatial regions within video frames and across time frames. Especially in scenarios with complex motion, occlusion, or background changes, it can capture cross-frame and cross-region contextual information, improving image coherence and detail restoration capabilities.
• Multi-head self-attention: Through multi-head self-attention technology, Vellum can further enhance the diverse expression ability of features and strengthen the ability to capture textures and details.

Multi-scale Feature Fusion

• Encoder-Decoder Architecture: Vellum employs an encoder-decoder architecture and combines it with a skip connection mechanism to effectively fuse features of different scales. The encoder performs downsampling layer by layer to extract abstract semantic information, while the decoder performs upsampling layer by layer to restore spatial resolution. Skip connections directly transfer high-resolution features from the encoding stage to the decoding end, effectively preventing the loss of spatial details.
• Multi-scale Fusion: Through multi-scale feature fusion, the robustness of the model to noise is enhanced, the accuracy of texture detail restoration is improved, making the enhanced video frames more natural and smooth, with rich details.

Reconstruction and Residual Learning

Residual Learning Framework: Vellum focuses on learning the residual information between the input video frames and the target high-quality frames, such as noise, blurring, and distortion, through residual learning. Residual connections accelerate the training convergence speed, avoid gradient vanishing, and ensure that the network can effectively focus on detail restoration and defect repair.

Local and Global Information Capture: By combining the depth of convolutional layers and skip connections, Vellum ensures that it can capture global structures while also restoring minute local details, ultimately achieving excellent image clarity and improved visual quality.

Training Objectives and Loss Function

Vellum's training employs a multi-task loss function, comprehensively optimizing multiple dimensions such as image reconstruction accuracy, edge sharpening, and texture preservation, to ensure the model's stable performance in different scenarios. Commonly used losses include: 

• Content loss (L1 or L2 loss): Ensures pixel-level similarity between the reconstructed image and the real image.
• Edge enhancement loss (such as gradient loss): enhances the clarity of image edges and detailed contours.
• Texture Preservation Loss (Perceptual Loss): Improves the realism and naturalness of textures, based on the feature differences of pre-trained convolutional networks.
• Adversarial Loss (GAN Loss): Reduces artifacts and enhances the naturalness and detail performance of images through generative adversarial training.

Application Scenarios and Performance

Vellum has powerful capabilities in multiple fields, especially when dealing with creatives that require high levels of detail and texture, it can provide excellent enhancement effects

• Architecture and Urban Planning
• Scenery and Natural Landscape
• Film and TV Production and Post-production

For a detailed model introduction, please refer to the relevant article 👉UniFab Texture Enhanced: Technical Analysis and Real-World Data Comparison

Kairo — Anime Optimization Model

Positioning: Kairo is an optimization model specifically designed for anime videos, featuring technical and architectural innovations tailored to the unique visual style of anime. It delivers high-quality upscaling and detail enhancement, preserving sharp lines and vibrant colors while balancing processing efficiency and visual coherence. Kairo accurately retains details in upscaled anime content and stays true to the original artistic style.

Core technological innovation of Vellum

Style-aware Convolutional Network

• Introducing a style-aware convolution module, which effectively enhances the clarity of line edges and the uniformity of color blocks by analyzing the unique edge lines and color block distributions in anime images. By learning the style characteristics of anime, Kairo ensures that the upscaled image adheres to the original artistic style, avoiding visual distortion caused by over-smoothing or over-sharpening. 

Color consistency maintenance mechanism

• To address the issue of color drift that often occurs in anime videos, Kairo integrates a color consistency maintenance mechanism, using spectral normalization and chromaticity separation strategies to ensure the color stability of the same character or scene across consecutive frames, avoid color differences or color discontinuities after upscaling, and enhance the viewing experience. 

Highly parallel lightweight architecture design

• To meet the requirements of multi-frame continuous processing for anime videos, Kairo employs techniques such as grouped convolution and tensor rearrangement to achieve model lightweight and high parallelization. This design not only significantly reduces computational resource consumption but also notably improves inference speed, meeting the real-time requirements for high-frame-rate animation upscaling.

Inter-frame consistency optimization

• One of the common issues in anime videos is inter-frame flicker and jump. Kairo effectively suppresses the visual jitter and detail incoherence that occur during the upscaling of consecutive frames by introducing an inter-frame consistency optimization technique in the time domain, ensuring a smoother and more natural viewing experience after animation upscaling.

Specialized edge protection strategy

• By combining edge-guided filtering and multi-scale edge detection strategies, Kairo can effectively enhance the extremely important line details in anime, ensuring that the lines after enlargement are sharp and free of jagged edges, thus avoiding the common problems of line blurring or breakage in traditional super-resolution methods. 

Multi-task Training and Loss Design

• During the training phase, Kairo combines content reconstruction loss, edge fidelity loss, and style preservation loss to ensure multi-dimensional optimization and high fidelity at the animation level. Through this multi-objective optimization, Kairo achieves a good balance among detail preservation, color restoration, and texture performance, ensuring that the visual effects of anime videos reach their best.

Kairo is a highly efficient enhancement model tailored for anime videos, focused on delivering high-quality upscaling and detail restoration. Utilizing style-aware convolutional networks, color consistency preservation, and inter-frame consistency optimization, Kairo maintains sharp lines and vibrant colors while ensuring upscaled frames remain true to the original artistic style. It is ideal for anime production, video editing, and animation rendering, offering an accurate and efficient solution for anime video upscaling.

For a detailed model introduction, please refer to the relevant article👉 UniFab Anime Model Iteration

Titanus — Flagship-level Video Enhancement Model

Positioning: Titanus is UniFab's flagship model specifically designed for high-resolution film and television creatives, aiming to provide ultimate image quality enhancement, especially suitable for movie-grade creatives and high dynamic range (HDR) content. Whether it is in movies, TV shows, documentaries, or other high-resolution videos in film and television production, Titanus can provide excellent detail restoration and image quality optimization, meeting the demanding requirements of professional film and television Post Production. This model will be released for the first time in UniFab 4, bringing more advanced technologies and features, and specific details will be explained in detail in subsequent series of articles.

Performance Comparison between UniFab and Topaz

4.1 Comparison of Image Quality Improvement Performance

From the perspective of visual enhancement performance, both use methods such as deep convolutional networks, attention modules, and adversarial training at the underlying level, but there are subtle differences in aspects such as detail recovery consistency, style fidelity, and temporal stability. 

Detail Restoration Ability

Topaz

• Proteus/Rhea excels in single-frame high-frequency texture restoration, especially suitable for compressed or mildly noisy scenarios.
• However, its general model occasionally exhibits inconsistent sharpening artifacts when dealing with anime and HDR movie creatives that do not conform to its training domain. 

UniFab

• Titanus enhances detail consistency in HDR, complex lighting and shadow, and film grain areas through multi-level convolution and dynamic compensation.
• Vellum's multi-scale ST-CNN can significantly improve scenes with high texture density, such as brick walls, forests, grasslands, etc., and reduce the "over-smoothing" problem commonly seen in traditional SR.
• Kairo can recognize the unique line structure and color block style of anime, avoiding line drawing breaks or color block banding artifacts. 

Function Coverage Comparison

Topaz: A modular system centered around visual enhancement tasks, Topaz Video AI covers the following core visual enhancement tasks:

• Enhancement
  • Proteus
  • Iris
  • Nyx
  • Rhea
  • Artemis
  • Gaia
  • Theia
• SDR to HDR
• Frame interpolation
• Stabilization
• Motion deblur

Its design centers around "visual restoration" but does not involve the complete video processing pipeline, such as transcoding, audio processing, or subtitle systems 

UniFab: A full-process audio and video processing system covering from input to output, UniFab integrates more than 18 audio and video modules internally, including:

AI Enhancement Category:

• Video Upscaler AI （Equinox/Titanus/Kairo/Vellum）
• HDR Upconverter AI（SDR→HDR）
• RTX Rapid Upscaler AI
• RTX RapidHDR AI
• Face Enhancer AI
• Denoiser AI (Silens)
• Smoother AI
• Video Colorizer AI
• Deinterlace AI
• Video Stabilizer AI

Video Processing Class:

• Video Converter
• Subtitle Generator AI
• Video Translator AI
• TV Show Converter
• Compress
• Video Background Remover AI
• Audio Upmix AI
• Vocal Remover AI

Compared to Topaz, UniFab is closer to an end-to-end video processing system, with stronger capabilities for integrating professional production processes.

Performance Comparison

In video super-resolution and enhancement tasks, processing speed is an important indicator for measuring system efficiency, especially in long video or batch processing scenarios. 

UniFab Processing Performance

Thanks to UniFab's in-depth optimization for mainstream GPUs (NVIDIA CUDA architecture), including: 

• Tensor Computational Graph Fusion
• Multithreaded parallel scheduling
• Video memory reuse and cache-friendly optimization
• FP16 Mixed Precision Inference

UniFab employs content type-based divide-and-conquer, with different models adopting different lightweight strategies to make its inference graph easier to accelerate. Meanwhile, UniFab has carried out deeper engineering optimizations at the framework level, including:

• No additional dynamic parameter adjustment is required during the inference phase
• The model calculation path is shorter and more stable
• High cache reuse rate
• More tightly coupled with modern GPU Tensor Cores 

UniFab has demonstrated significant inference efficiency advantages in multiple actual test scenarios, with an average of:

Video enhancement speed of approximately 8 - 10 Frames Per Second (FPS)

This means that in high-load tasks such as 1080p→4K and 4K→8K, UniFab can provide lower processing latency and higher throughput . 

Topaz Performance

Under the same GPU configuration, the processing speed of Topaz Video AI usually remains at:

Approximately 3 - 5 FPS

Due to its relatively larger model size, higher number of parameters, and greater memory footprint, it is more susceptible to the impact of memory
bandwidth and computing power bottlenecks during the inference process.

Price Comparison

Future Outlook and Upgrade Plan

Technical Optimization Direction

UniFab plans to continue optimizing its algorithms, especially in enhancing detail recovery and depth enhancement. In the future, UniFab will place greater emphasis on AI self-learning and model Self-Adaptation capabilities to provide more precise processing solutions for different types of creatives.

Expected New Models and Application Expansion

UniFab also plans to launch more models specifically tailored to specific scenarios (such as slow-motion video enhancement, facial detail restoration, etc.) in the future.

👉 Community Link:  UniFab 4.0 is coming soon!

Welcome you to share topics or frame interpolation models that interest you on our forum. We will regularly publish technical reviews and version updates to drive continuous improvement and carefully consider your feedback from testing and evaluation.

Preview of the next article: UniFab NEW Upscaler Model —— Titanus

Previous:

📕 UniFab Anime Model Iteration

📗 New Features | UniFab RTX RapidHDR AI Features in Detail

📘 The Iterations of UniFab Face Enhancer AI

📙 UniFab Texture Enhanced: Technical Analysis and Real-World Data Comparison