Conclusion

I came away from NVIDIA's briefing considerably more impressed than I expected to be. All of their decisions make sense, and most of them are what I would have done in their place, too. Conditioning the model on a rendered frame instead of a prompt. Training it so that it can't affect geometry. Running diffusion in full pixel space. Taking control away from the model and putting it in the hands of the developers. None of those are the "easy" solutions, and the visual outcome looks good.

What It Actually Adds
The NBA 2K27 comparisons clearly show that, and it's not "dramatic." Faces are not "redrawn" like copy and paste from Google images. Rather there are small nuanced improvements—occlusion in an eye socket, a contact shadow where a shoe meets the floor, the shadow cloth casts on the skin underneath it. That's precise work, adding features that are very expensive to compute otherwise.

The Price
50%+ of your frame rate is an enormous amount to pay, especially if it's on top of ray tracing. NVIDIA's own numbers only look high because Multi Frame Generation 6X is used: the RTX 5090 gets 370 FPS at 4K, which really is a DLSS-upscaled 1080p render, multiplied six ways—and it still looks good. On demanding games, I'm not sure how that's gonna work out. What gives me hope is that currently the model is FP8, which means an NVFP4 version could boost FPS significantly.

RTX 50 Only
According to NVIDIA, DLSS 5 is GeForce 50-series only, but the modders have shown that it doesn't have to be. Within five days, without the thousands of engineers that NVIDIA has, the community had DLSS 5 running on Ada, on Ampere, on a 2019 RTX 2060, and in games and APIs NVIDIA has never mentioned. This demonstrates that the limitation is software, timings, probably performance, too. I expect official support to broaden, and that's also why I expect player-facing controls eventually. The reason is that there will be never-ending drama about it, resulting in an infinite stream of bad PR for NVIDIA.

Improvements
DLSS 5 sits between Super Resolution and Frame Generation, both of which hold a great deal of temporal information about the frame it is working on, and it is currently blocked from using any of it to correct anything. A generative model that understands a scene semantically offers unique capabilities that none of the other technologies in the DLSS stack bring, and I feel that this could be a gold mine for reconstruction filters, to resolve ghosting, smearing, shimmer and other artifacts.

Game Adoption
Adding DLSS 5 costs a studio almost nothing: the engine already generates the required inputs, so the real work is a toggle in the Options menu, at least to have a baseline checkmark "DLSS 5 Supported." Tuning it properly will be much more expensive, and I'm not convinced gamers will riot and demand it, considering we're seeing titles ship with the most basic bugs around DLSS, like wrong profile names, wrong requirements etc. I still expect DLSS 5 to spread quickly and widely, and I'm slightly worried whether it will become the new default, which lets studios skip on the underlying quality in the way they build their characters especially. Motion capture and high-poly models are expensive, just let AI yolo it?

Artistic Intent
NVIDIA found a solid solution here. The developers are the only ones controlling DLSS 5, with various sliders and settings, so anything they ship is by definition the artist's intent. Whether that's enough to ensure the blame never lands on NVIDIA, I doubt it, they built the model after all. Handing the decision to the studio is still the right design, but I hope they get it right, and great results will likely be on a case-by-case basis.

AMD Falls Further Behind
It is hard to look at this and not think about AMD. They spent the last two years closing the gap on upscaling, and FSR 4 finally put them in the same ballpark as DLSS on image quality. In the meantime, NVIDIA opened an entirely new frontier. There is nothing similar from AMD or from Intel that we know of, and the barrier to entry is not a shader technique anymore. It is a trained generative model, which requires lots of compute power, and time, and input data, and tensor hardware fast enough to run it inside a single frame.

Image Quality
What we have seen so far looks extremely promising. But we don't know yet whether an entire game rendered this way works just as well. The modders have shown excellent results and people in forum threads and on Reddit are pretty impressed. One thing that's worth mentioning is that some faces seem to look older now, possibly by the added detail and the small shadows registering as "older." It happens to young characters too, which is what makes it a pattern rather than a coincidence. Some of that is surely the maximum-uplift, nothing-masked configuration the mod scene runs at, and it is exactly why we need tuned settings.

Exciting Times Ahead
There will always be people who reject all the new tech. That is a legitimate position, and it is not going to change. But the technology underneath is genuinely new, think of DLSS 1. It will improve constantly, and the modding community will keep pushing NVIDIA to improve it—and wherever NVIDIA will not, they will simply deliver it to the community themselves.