AMD's FSR MLFG vs. FSR 3: Quality Gains and Frame Pacing Problems

Machine learning frame generation introduces notable visual improvements while retaining the long-standing pacing inconsistencies observed in earlier FSR versions.

Hardware by Masaru Hoshino on Dec 19, 2025

Evasar Redstone is launched—well, kind of. One of its core features, the Radiance Cache, is only coming to a future title, Warhammer 40k Darktide. Ray regeneration, another feature, is currently only in one title, Call of Duty: Black Ops 7's multiplayer component.

That leaves the FSR Frame Generation update, which now uses machine learning on RDNA 4 GPUs and can be applied to several games via an in-driver panel update, similar to how FSR 4 Super Resolution launched for many titles at the beginning of last year.

AMD, FSR MLFG, FSR 3, Quality Gains, Frame Pacing Problems, NoobFeed

For FSR 4 Super Resolution, the Redstone launch did not cover support for that feature on older GPUs, even though versions of it have been seen running on RDNA 2 and RDNA 3 through community efforts. Redstone's launch is therefore a strange half-launch. That strangeness extends, to some degree, to the newly debuted machine learning upgrade for Frame Generation.

Quality Improvements in Machine Learning Frame Generation

As other outlets and Hardware Unboxed have reported, machine learning-based frame generation from AMD delivers significant quality gains over the previous technique. A clear example appears in Cyberpunk 2077. When FSR 3 Frame Generation is enabled, the system uses motion vectors from the game to reproject visuals between frames, making motion appear more fluid.

But it doesn't have strong data for things like shadows. If you look closely at the shadow behind the automobile in the benchmark sequence, you can notice that it falls behind and often shows the wrong perspective. DLSS 3 and DLSS 4 Frame Generation historically outperform FSR 3 in these cases, producing fewer artifacts.

Comparing FSR 3 to FSR MLFG shows a similar quality upgrade. The new MLFG uses a more intelligent, automated method to combine optical flow data from world features, such as shadows—which lack motion vectors—with optical flow from objects that do have motion vectors.

As a result, MLFG produces more convincing movement, resembling what a traditional non-generated frame might look like. The same applies to particles. In FSR 3, particles show ghosting and appear low-frame-rate when mixing generated and traditional frames. FSR MLFG presents particles in a way much closer to expected visual continuity.

While not entirely at the quality level of DLSS 4 FG, FSR MLFG is significantly better than past results. Its experiential quality approaches what you might find on an RTX GPU. In many scenarios, the difference becomes hard to detect. Flashing traditional frames with generated ones is forgiving, which helps MLFG appear convincing where FSR 3 was noticeably inferior.

AMD, FSR MLFG, FSR 3, Quality Gains, Frame Pacing Problems, NoobFeed

Persistent Issues With Frame Pacing

Unfortunately, this quality upgrade carries over long-standing frame-pacing issues seen in FSR 3. These concerns have already been heavily covered, particularly by Hardware Unboxed. The core issue shows clearly in Industria 2, a worst-case scenario for AMD frame generation.

This title helps MLFG by letting you upgrade your drivers and adding the in-game FSR option without having to switch between frame generation options. It makes frame pacing very unstable, even though the average frame rate is over 120 fps.

The output doesn't have a steady pattern; instead of always having 7–8ms frame times, it jumps between other values like 4ms, 8ms, 12ms, 16ms, and more. The leaps from 8ms to 12ms already make things less fluid, but the bigger jumps from 4ms to 12ms, 8ms to 16ms, or 4ms to 16ms make motion quite bad.

This makes the game appear to stop and resume on a VRR display, where it seems to move for a short time before staying static for longer than intended. The effect is visually unpleasant and goes against the goal of frame generation.

The technology also produces runt frames, which are scanned out so quickly that they appear only as fleeting tears on the screen. These frequently occur within 0-1 ms, which is below the display's maximum refresh rate. This makes them look like they are tearing, which is the only way to see them.

Turning off Frame Generation in the same title on the same GPU makes the output much smoother. The game seems much smoother on a VRR panel, even if the average frame rate is lower and the overall frame duration is longer. This is because there is no annoying frame time fluctuation.

Not all titles have this problem. Cyberpunk 2077 and Black Myth: Wukong have superior pacing overall, with several parts that are clearly fluid. But even there, there are times when things aren't stable, which can make some parts of the game run poorly.

Additional Examples of Frame Pacing Instability

God of War Ragnarok is another example of frame time volatility that isn't good. When you look at a rock, the frame pacing goes up and down a little, but not enough to be very noticeable on a VRR panel. But when you turn the camera, it adds runt frames that are less than 3ms long and creates the same stop-and-go action that you see in other games. These runt frames mostly appear as quick tears, making it hard to show the movement intended.

These inconsistencies continue from one title to the next and even from one moment to the next in the same game. The inability to guarantee stable frame pacing makes it difficult for players to trust that enabling FSR Frame Generation will consistently improve the experience.

The competition manages this aspect significantly better at present, which highlights the importance of AMD addressing this weakness.

AMD, FSR MLFG, FSR 3, Quality Gains, Frame Pacing Problems, NoobFeed

Why the Discussion Has Re-Emerged

Although this issue has been known since FSR 3's launch, its return to discussion is due to the new MLFG release and differing data presentation across outlets. Frame rate graphs do not tell the story because they fail to reveal frame-pacing problems. Frame time graphs reveal issues more clearly, but even then, methodology matters.

Some places use PresentMon, which can show frame durations based on the MS between presents. This is an older standard that might hide problems with pacing. More precise internal methods, such as MS between display change, show spikes and runt frames.

FCAT analysis goes even further by using color-bar tagging and post-processing on a second PC to measure the actual durations of scanned-out frames. Even though it's a bit of a pain, this method is quite accurate and nearly matches what you see.

Tim's testing with cameras is even more realistic because it shows how an uneven pace affects things in ways that graphs or metrics can't.

Final Thoughts

FSR MLFG delivers significantly improved frame quality, closer to industry standards. But frame pacing is still not uniform across games and titles, which means VRR displays can't be sure to remain smooth.

Since smoothness is a key goal of frame generation, AMD should aim to maintain a continuous pace similar to that of other solutions.

The capability for smooth output clearly exists, and users deserve the reliability that should accompany frame generation technology.

Also, check our other AMD articles below: