12 Minutes of Starfield Gameplay with DLSS 5: NVIDIA Promised Magic, But We Saw Weird Shadow Dancing

A few days after the unveiling of DLSS 5 at GTC 2026, which was met with a wave of criticism, the first in-depth look at how the new technology performs in real-world conditions has surfaced online. Blogger Plano Plays Games published a screen recording on their channel featuring a 12-minute demo of Starfield with the new upscaling algorithm both on and off.

Despite the mediocre quality of the recording itself, the difference in visual fidelity is noticeable to the naked eye. When DLSS 5 is activated, scenes start to look significantly more detailed: proper lighting appears, shadows are refined, and objects no longer look flat. The effect is especially evident in vegetation—foliage and branches gain a sense of volume, and that feeling of looking at a texture stretched over polygons disappears.

However, artifacts are also present, and the recording clearly shows strange shadow “dancing” or jitter when the character moves sideways. Additionally, when transitioning from dark interiors to well-lit spaces, lighting effects can be seen loading in almost in real-time. Also noteworthy is the complete absence of active scenes involving gunfights or dynamic movement in the video. Only a slow, leisurely walk through locations is shown; the demo creators deliberately avoided scenarios where the algorithm might not perform at its best.

Regarding the technical details gleaned from the video, it’s worth noting that DLSS 5 operates on a transformer model and can be integrated into games without native ray tracing support. The technology analyzes the color buffer and motion vectors, after which the neural network “fills in” the lighting, reflections, and shadows. According to NVIDIA representatives, integration only requires DLSS 4 support—developers don’t need to write separate code. However, this approach has raised questions among observers: the algorithm essentially doesn’t know what’s actually happening in the frame, and works solely with pixels.

The release of DLSS 5 is scheduled for the fall of this year, and only then will it be possible to finally assess how ready the technology is for widespread use. For now, the public has to make do with these leaked snippets of information, which leave more questions than they provide answers.

Are you ready to have a neural network paint in the lighting in your games, or would you prefer classic ray tracing with a clear, understandable logic behind how it works, even if it takes a bigger hit on performance? Share your thoughts in the comments.