ML-Based Frame Generation
When rival NVIDIA introduced DLSS 3 Frame Generation, in which the GPU attempted to draw every alternate frame entirely using AI to double frame-rates, AMD responded with FSR 3 Frame Generation, which used more primitive frame interpolation techniques. AMD has finally caught up with image quality with its new FSR Redstone ML-powered Frame Generation.
The new AMD FSR Redstone ML-powered Framegen uses a combination of optical flow processing and motion vector re-projection to generate alternate frames. The model works by taking inputs of the previous and current frames, and their derivative depth and motion vectors to help it guess what the next frame could look like.
There are two machine learning models in use, one drives the optical flow reprojection, and the other predicts color at the blend stage when it is combining inputs from the optical flow and motion vectors.
The frames generated by FSR Redstone tend to have significantly higher image quality than the interpolated frames generated in FSR 3 Frame Generation. Perhaps the best evidence of this is in the F1 25 screenshot where the shadows of the cars come out more accurately.
As we stated earlier, ML-based frame generation can either be integrated with the game and appear in the game settings, or it can be enabled through a driver-level override.
Supported Games
At this time, the following games support ML Frame Generation via driver-level override on RDNA 4 cards:
- ARC Raiders
- Black Myth: Wukong
- Call of Duty: Black Ops 7
- Cash Cleaner Simulator
- Cronos: The New Dawn
- Cyberpunk 2077
- Deathground
- EXFIL
- F1 25
- Farming Simulator 25
- God of War Ragnarok
- Grand Theft Auto V Enhanced
- Hell is Us
- Hogwarts Legacy
- INDUSTRIA 2 (demo)
- Keeper
- Mafia: The Old Country
- Marvel Rivals
- No More Room in Hell 2
- Project Motor Racing
- Qanga
- Rem Survival
- Ships At Sea
- Squad
- THE FINALS
- The Last Caretaker
- The Talos Principle: Reawakened
- Titan Quest II
- Winter Survival
- Wreckfest 2
- Wuthering Waves