Neural Temporal Supersampling for TAA Ghosting Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Temporal Anti-aliasing (TAA) techniques suffer from ghosting artifacts around moving object boundaries due to frame-to-frame changes in visibility and errors in motion vectors, leading to lower quality upsampling when rendering at native resolution.

Innovation Solution

A mixed low precision convolutional neural network is used for temporally amortized supersampling, incorporating sample reliability information to improve image quality and reduce blurring, while rendering at lower spatial resolution for a performance boost.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If Temporal Anti-aliasing (TAA) accumulates temporally supersampled images, then anti-aliasing quality is improved, but ghosting artifacts appear around moving object boundaries

Engineering Contradiction:
Improveanti-aliasing qualityVSAvoidghosting artifacts
Core Design Contradiction:
Manufacturing precisionVSObject-generated harmful factors

Solution Approach 1:

The system performs preliminary actions by accumulating temporally supersampled images before final rendering, and also discards outdated frame data in advance to prevent ghosting artifacts from forming in the first place

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system changes parameters by adjusting the accumulation window size and temporal weighting factors dynamically, allowing it to optimize between anti-aliasing quality and ghosting reduction based on scene motion characteristics

Inventive Principle:
Principle #35Parameter changes

2Stability of the object's composition

If TAA warps previous frame using motion vectors, then temporal alignment is improved, but mismatches occur due to motion vector errors

Engineering Contradiction:
Improvetemporal alignmentVSAvoidsample matching accuracy
Core Design Contradiction:
Stability of the object's compositionVSManufacturing precision

Solution Approach 1:

The system uses feedback by comparing accumulated samples with current frame data and adjusting temporal weighting accordingly, allowing it to detect and correct mismatches caused by motion vector errors

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system applies partial temporal accumulation rather than full frame warping, using only the necessary portion of historical data to maintain alignment while avoiding the propagation of motion vector errors

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If rendering is performed at native resolution, then image quality is improved, but render time increases

Engineering Contradiction:
Improveimage qualityVSAvoidrender time
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The system transitions from purely spatial rendering to temporal-spatial supersampling, accumulating samples across multiple frames to achieve native resolution quality at lower effective render rates

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system performs preliminary supersampling accumulation over multiple frames before final output, distributing the computational workload over time to reduce peak render time while maintaining quality

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4138028B1Sample distribution-informed denoising & rendering
Publication Date: 2025.12.10 INTEL CORP
  • EP4138028B1 patent drawingFigure 1
  • EP4138028B1 patent drawingFigure 2A
  • EP4138028B1 patent drawingFigure 2B

AI summary

A graphics processor is provided that includes circuitry configured to receive, at an input block of a neural network model, a set of data including previous frame data, current frame data, velocity data, and jitter offset data. The neural network model is configured to generate a denoised, supersampled, and anti-aliased output image based on reliability metrics computed based on sample distribution data for samples within the current frame data.