Browsed by
Category: AI

Tool for measuring AI enhanced GPU image quality

Tool for measuring AI enhanced GPU image quality

Engineers at Intel released an open-source tool that tries to quantify the issues from increasing amounts of upscalers, frame generators, and AI rendering techniques. Ironically, the tool itself is an AI trained on large datasets. Their paper about the methodology is located here.

CGVQM is a video quality metric that predicts perceptual differences between pairs of videos.
Like PSNR and SSIM, it compares a ground-truth reference to a distorted version (e.g. blurry, noisy, aliased).

What sets CGVQM apart is that it is the first metric calibrated for distortions from advanced rendering techniques, accounting for both spatial and temporal artifacts.

CGVQM is available for free on github and uses PyTorch optimized for CUDA GPUs though it does work on CPUs.

Other links:

AI GPU/Video Quality Metrics

AI GPU/Video Quality Metrics

How do you know when AI generated frames on your graphics card or in video playback look bad? Some folks at Intel published a paper and shared a dataset for a technique called CGVQM+D: Computer Graphics Video Quality Metric and Dataset – which tries to quantify the image quality issues created by using AI generation techniques like neural supersampling, novel-view synthesis, path tracing, neural denoising, frame interpolation, and variable rate shading.

AI to AI communication

AI to AI communication

It’s no secret AI agents are starting to take over customer service and other tasks. So, instead of communicating via expensive text-to-speech conversion, why not use their own language.

Gibberlink allows AIs to skip unnecessary speech synthesis and recognition and communicate using beeps and bleeps like R2D2.

Protect yourself from Skynet

Protect yourself from Skynet

Drone warfare is very real as the conflict in Ukraine has shown the world. Cheap, re-purposed drone airplanes deliver bombs behind enemy lines and small copters are being used to deliver decisive and deadly grenades to entrenched positions. Hoog created this distopian instructional video that imagines how bad things could get and how to protect yourself from packs of rogue robots.

While this version is tongue-in-cheek, the reality is that swarms of cheap drones can absolutely overwhelm and kill larger, well-equipped enemy forces.

Real anti-drone weapons are already being developed

Welding class

Welding class

When I was on a 8 week sabbatical a number of years ago, I took an artistic welding class. It was a ton of fun. Sugar Metal Customs does custom ironwork, but also teaches some classes from making metal roses to giant metal flails for your medieval fantasies.

Maybe pick up a skill that AI can’t replace?

AMD accidently releases FSR

AMD accidently releases FSR

AMD recently published a new version of its FidelitySDK with FSR 4 upscaling and FSR 3.1.5 frame generation support. Unfortunately, they accidentally published the full FSR 4 source code on GitHub. Before AMD took it down, some media outlets and X users managed to record screenshots of the files, including Videocardz.

Update: PCGamer is reporting that AMD is having more trouble putting the genie back in the bottle than first thought because they accidently included the MIT license.