Edited by 
Short answer
DLSS (Deep Learning Super Sampling) is NVIDIA's set of AI features for games that run on the Tensor Cores of RTX graphics cards. Super Resolution renders each frame at a lower resolution and uses a neural network to upscale it. Frame Generation creates extra frames between rendered ones. Ray Reconstruction replaces hand-tuned denoisers for ray-traced lighting. The current version, DLSS 4.5 from January 2026, uses a second-generation transformer model for upscaling on all RTX cards and adds up to 6x multi frame generation on the RTX 50 series.
DLSS features and which cards get them
| Feature | What it does | RTX 20 / 30 | RTX 40 | RTX 50 |
|---|---|---|---|---|
| Super Resolution | AI upscaling from a lower render resolution | Yes | Yes | Yes |
| DLAA | The same AI model used for anti-aliasing at native resolution | Yes | Yes | Yes |
| Ray Reconstruction | AI denoising for ray-traced and path-traced lighting | Yes | Yes | Yes |
| Frame Generation | One generated frame for each rendered frame | No | Yes | Yes |
| Multi Frame Generation | Up to 3 generated frames per rendered frame (4x), and up to 6x with DLSS 4.5 | No | No | Yes |
How DLSS works
Super Resolution
The game renders at, for example, 1080p instead of 4K, which needs about a quarter of the pixel work. DLSS takes that frame plus motion vectors, depth and the previous output, and a neural network running on the Tensor Cores predicts the 4K image. Using information from earlier frames lets it rebuild detail that a single low-resolution frame does not contain. The CUDA cores are freed up and the frame rate rises, often by 2x or more in demanding games.
Frame Generation
Frame generation predicts entirely new frames between two rendered ones. It raises smoothness but not responsiveness: your input still updates at the rendered frame rate, and generating frames adds a little delay. That is why NVIDIA pairs it with Reflex, which cuts latency elsewhere in the pipeline, and why frame generation works best when the base frame rate is already reasonable.
Ray Reconstruction
Real-time ray tracing can only afford a few rays per pixel, which leaves noise. Games traditionally cleaned it up with hand-written denoisers. Ray Reconstruction replaces them with one AI model trained on high-quality images, which keeps more detail in reflections and lighting. See Tensor Cores vs RT Cores for how the two units share this work.
Every DLSS version
| Version | Year | What changed |
|---|---|---|
| DLSS 1 | 2019 | First AI upscaler, trained per game; often blurry |
| DLSS 2 | 2020 | One general model using motion vectors and past frames; a big jump in quality |
| DLSS 3 | 2022 | Frame Generation, for RTX 40 cards |
| DLSS 3.5 | 2023 | Ray Reconstruction, for all RTX cards |
| DLSS 4 | 2025 | Transformer models instead of convolutional networks; Multi Frame Generation up to 4x on RTX 50 |
| DLSS 4.5 | 2026 | Second-generation transformer for Super Resolution, said by NVIDIA to need 5x the compute; dynamic multi frame generation up to 6x on RTX 50 |
Why Tensor Cores and FP8 matter
Every DLSS feature is a neural network that has to run in a few milliseconds per frame, which is only possible on Tensor Cores. Newer models are heavier: NVIDIA says the DLSS 4.5 upscaling model needs about five times the compute of the first transformer model. RTX 40 cards run it in FP8 and RTX 50 cards add FP4, while RTX 20 and 30 cards lack FP8 Tensor Cores, so they take a larger performance hit from the new model. Older cards can still choose the earlier models in the NVIDIA app. Read FP16, BF16, FP8 and FP4 explained for why smaller formats run faster.
DLSS vs FSR and XeSS
- AMD FSR: versions up to FSR 3 use hand-written algorithms that run on almost any GPU. FSR 4 uses machine learning and runs on Radeon RX 9000 cards.
- Intel XeSS: an AI upscaler that runs best on Intel Arc GPUs and has a fallback mode for other cards.
- DLSS: runs only on NVIDIA RTX cards and is supported in more games than either.
Frequently asked questions
Does DLSS reduce image quality?
At its Quality setting, modern DLSS often looks as good as native resolution, and sometimes sharper thanks to better anti-aliasing. Lower settings such as Performance trade some detail for higher frame rates. DLAA uses the same model at native resolution purely for image quality.
Is DLSS frame generation "fake frames"?
Generated frames are predicted rather than rendered, so they make motion smoother without making the game respond faster. They work best when the game already runs at 50 to 60 frames per second or more before frame generation.
Can GTX cards use DLSS?
No. DLSS needs Tensor Cores, which GeForce cards have only from the RTX 20 series onward.
How do I use DLSS 4.5?
In games that support DLSS, the NVIDIA app can override the model to the latest version, so a game does not need an update to use the new upscaler.
