AMD FSR Redstone in detail – milestone and real counterpoint to NVIDIA’s DLSS?

The next scene under rainy conditions again shows the characteristic further development of FSR4 compared to FSR3, albeit in a visually much more complex environment. The combination of a wet surface, small-scale reflections, fine geometry and atmospheric scattered light poses challenges for any upscaling process, especially when the native output resolution is reduced. In the FSR3 display, the scene appears somewhat softer overall. The lack of precision is particularly noticeable in smaller structures such as the fences, the translucent rain tracks and the reflective surfaces of the vehicles. These structures are sometimes merged or lose their defined edge detail. The fine details of the wet ground, such as reflections in puddles or the structure of the debris, appear inconsistent and look washed out in places. Objects in the background, especially the lighting elements and the screen on the billboard, also show reduced clarity and slight double contours.

FSR4 shows a visibly clearer reconstruction in direct comparison. The vertical contours of the light poles are more stable, rain streaks are reproduced more finely and appear less “blocky”. Reflections on car roofs and wet surfaces have a more uniform structure. The finely stratified latticework in the background is also more clearly recognizable. The overall sharpness is higher, without this leading to artificial oversharpening artifacts. Small geometries, such as signs, crash barriers and fencing, appear more consistently segmented and lose fewer fine details.

Despite these advances, certain limitations remain visible. Lettering, fine textures and small-scale patterns appear clearer under FSR4 than under FSR3, but do not achieve the precision one would expect from NVIDIA’s DLSS. Due to its strongly data-driven model, DLSS tends to fill in missing structures with plausible patterns, which in many cases leads to an impressive reconstruction. AMD’s ML model is more conservative and hardly adds any additional information. While this reduces image errors, it also limits the ability to reconstruct very small text or complex detailed structures convincingly.

In these scenes, FSR4 shows significant improvements over FSR3 in terms of sharpness, stability of fine structures and coherence of reflective surfaces. Spatial details are better preserved, rain and lighting effects are reconstructed more homogeneously and the overall impression is more natural. Nevertheless, it is clear that AMD remains more reserved with FSR4 in Balanced mode when it comes to restoring very fine textures. In contrast to DLSS, which often adds missing structures due to the training process, FSR4 primarily reproduces the existing input signal and largely dispenses with additional interpretations.

It should also be noted that still images are only of limited informative value. One of the main problems of earlier FSR versions was temporal artifacts during fast movements, for example in the form of streaks, incorrectly reconstructed edges or partial loss of content. FSR4 reduces these effects, but does not yet eliminate them completely. Especially in dynamic scenes with fast rotations or changes in viewing direction, it is crucial how stable the temporal component really works. The improvements shown here in static individual images are therefore a clear step forward, but are no substitute for the evaluation in real game situations. These have improved considerably, but it will probably also depend on future implementations.

All in all, FSR4 offers a visibly improved reconstruction compared to FSR3, but still lags slightly behind DLSS in terms of detail reproduction and temporal stability. The development is definitely going in the right direction, but the technological gap to NVIDIA’s mature ML pipeline still exists in some areas, especially with fine textures, small-scale typography and fast image sequences.

Summary and conclusion

The overall picture that emerges from all the analyses is one of the clearest advances AMD has made since the introduction of FSR. With Redstone, ML upscaling, Ray Regeneration and the newly launched Frame Generation, AMD is for the first time moving into a technological framework that is no longer based solely on algorithmic approaches, but on real learning-based improvements in temporal and spatial reconstruction. This step fundamentally changes the classification of FSR and brings the entire ecosystem much closer to modern AI render pipelines, as NVIDIA has already successfully established for several generations.

The performance analyses show that AMD’s approach not only works, but also works extremely efficiently in many cases. Especially in combination with frame generation, significant performance gains are achieved, which are remarkably favorable in relation to the GPU performance consumed. The efficiency per frame increases significantly in ML modes because synthetic intermediate images only require a fraction of the energy of a fully calculated frame. FSR4 utilizes this fact more consistently and homogeneously than FSR3, which is reflected in more stable frametimes and better variance values. Overall, the pipeline appears more structured, less susceptible to temporal dropouts and capable of generating a technically comprehensible performance picture even in complex scenes.

At the same time, the comparisons show that FSR4 clearly improves image quality. Structures are reproduced more coherently, edges are reconstructed more cleanly and fine details are captured better than with FSR3. The ML models produce a smoother overall image, especially in scenes with small-scale geometry or diffuse structures such as rain, traffic lighting or complex shadow patterns. The number of temporal artifacts decreases, even if they do not disappear completely. Especially in quiet scenes, FSR4 reveals that the purely heuristic limitations of the older FSR generations have been overcome.

However, AMD’s ML approach remains deliberately conservative. Unlike NVIDIA’s DLSS, the model makes little attempt to interpret missing information or add details. This approach has advantages, as it reduces the risk of noticeable misreconstructions and preserves the structural truth of the scene, but it also limits the ability to fully restore extremely fine typography or highly detailed textures. DLSS still has the edge here because the model is more data-driven and is more often capable of making plausible additions to the image structure. This remains a factor for demanding users who are looking for perfect detail reproduction.

The temporal component is also not yet fully at the level of NVIDIA. Although many of FSR’s previous weaknesses have been reduced, fast rotations, sudden changes in direction or very dynamic scenes can still produce slight artifacts. The number of problematic image areas is significantly smaller than in FSR3, but FSR4 still does not quite achieve absolute stability – the kind of visual stability that DLSS 3.7 or higher works with in many engines. Although the differences become smaller and smaller when playing, they remain visible.

In the end, the overall impression is very positive. FSR4 is a big step forward that takes AMD into a new generation in almost all relevant disciplines. The performance is strong, the efficiency impressive and the image quality noticeably improved. The new ML pipeline provides a technically advanced foundation that will also enable the introduction of more advanced technologies such as Radiance Caching in the long term. For gamers who use a Radeon GPU or prefer an open, broadly integrable upscaling technology, Redstone is the first alternative that scores points not only for compatibility and openness, but also for technical substance.

The gap to DLSS is clearly shrinking, but not disappearing. NVIDIA still has a slight advantage in the areas of temporal stability, detail reconstruction and training scope, especially with very fine surface structures and in highly dynamic scenes. But compared to previous FSR generations, the progress is considerable. FSR4 not only offers a visible improvement, but also lays the foundation for further development steps that will be decisive in the coming months and years.

In the final section, it can therefore be concluded that with FSR Redstone, AMD is delivering an upscaling system for the first time that can no longer be considered just a necessary alternative, but a serious competitor. The technical basis is right, the performance gains are real and the image quality has reached a level that is very convincing in many games. The solution is not yet perfect, but the path is clearly recognizable and is clearly leading the way forward.