What do CNN and transformer mean?
The two words describe how an AI model is built. You do not need a computer science degree to understand the difference, because an everyday comparison works well.
The CNN approach: looking at small neighbourhoods
A convolutional neural network, or CNN, studies an image through small windows that slide across it. Each window sees only a small neighbourhood of pixels at a time. It is like inspecting a painting with a magnifying glass, one patch after another, and combining what each patch shows. This approach has served image processing well for many years, and it powered DLSS from the second version onwards.
Strengths of the CNN design
CNN models are efficient, because they reuse the same small windows everywhere. That keeps the cost low and made them a good match for real-time games.
Limits of the CNN design
Because each window sees little context, CNNs can struggle when a decision needs information from far across the picture. In games this showed up as ghosting behind moving objects, shimmering on thin lines and some loss of fine detail.
The transformer approach: looking at relationships across the picture
A transformer model uses a mechanism called attention, which lets it weigh how different parts of the input relate to each other, even when they are far apart. It is like stepping back from the painting to see how the whole composition fits together, while still noticing the brushwork. Transformers became famous through language tools, and the same idea applies well to images and video.
Why this helps upscaling
To rebuild a sharp frame, the model must decide which details from earlier frames still apply and which are out of date. A model that sees broader context can make better decisions, which reduces ghosting and keeps detail stable.
What it costs
Transformer models are larger and take more calculation. Modern RTX cards have enough Tensor core power to handle it, but the cost is real, a point covered below.
Where DLSS used each model
The history clarifies which games you may meet with which model.
DLSS before version 4
Versions 2 and 3 used CNN models for Super Resolution and Ray Reconstruction. Many games still ship with these older models, because developers update DLSS files only when they release a patch.
DLSS 4 and later
DLSS 4, introduced with the RTX 50 series in early 2025, brought transformer models to Super Resolution and Ray Reconstruction. NVIDIA announced DLSS 4.5 at CES 2026 with a second-generation transformer model for Super Resolution. Our overview of DLSS 4 and the whole family places each step on a timeline.
What stays the same
DLAA, which applies DLSS-quality image processing at native resolution, can use the same model families. The article on DLAA explains when to choose it.
What improved with the transformer model
NVIDIA's claims and independent comparisons point to the same areas.
Stability
Edges and fine patterns stay steadier from frame to frame, with less crawling and flicker. This is the change most players notice first, in foliage, fences and distant buildings.
Detail
Fine detail such as text on signs, hair strands and thin wires often looks cleaner, particularly in the Performance modes where the starting image is small.
Motion clarity
Moving objects leave fewer ghost trails. This helps in racing games, shooters and any scene with fast camera movement.
Better ray-traced lighting
The transformer version of Ray Reconstruction produces clearer lighting with less smearing in ray-traced scenes. The guide to Ray Reconstruction and supported games explains this separately.
What did not change
Upscaling still cannot create detail that was never rendered. At low internal resolutions the picture is still softer than native. The transformer model improves the rebuild, not the laws of information.
What does it cost in performance?
The transformer model asks the card to do more work for each frame, so frame rates can dip slightly compared with the CNN model at the same mode.
Newer cards handle it well
NVIDIA has described the cost on RTX 40 and RTX 50 series cards as a minor trade-off, and independent tests have generally found small differences of a few percent in many games. Exact results depend on the game, the resolution and the mode.
Older cards pay more
RTX 20 and RTX 30 series cards lack some of the hardware features that newer models use, so the same model can cost them noticeably more. Reports on the second-generation transformer model in DLSS 4.5 found larger drops on older cards than on newer ones. If frame rate is tight on an older card, the first-generation transformer model is usually the safer choice.
How to measure on your computer
Run a repeatable scene with the CNN model, then with the transformer model, and compare average frame rate and the slowest frames. The site's guide to benchmarking tools and methods explains a fair test.
How to switch models
There are two routes, and neither is difficult.
In the game
Games that ship with DLSS 4 or later usually use the transformer model already. Some games offer a choice in their menu, often labelled by model or quality. Check the graphics settings, and read the game's patch notes for DLSS updates.
With the NVIDIA App
For games that use an older DLSS, the NVIDIA App can override the model. Open Graphics, then Program settings, choose the game and open Driver settings. Look for the DLSS override for model presets and choose the latest option. The article on DLSS presets explains the lettered names, and the guide to using the NVIDIA App shows the screens.
Turn on the on-screen indicator
NVIDIA's overlay can show which DLSS version and preset a game is using, which helps you confirm that an override took effect.
How to undo it
Set the override back to the game's own setting. Nothing permanent changes on your computer.
Should you switch?
For most players, yes. The reasons are simple.
Reasons to switch
The picture is steadier and cleaner in most games, the cost on recent cards is small, and the switch is reversible. If you are using Performance or Balanced mode, the gain is often larger, because the AI has less to work with and benefits more from a better model.
Reasons to wait
If you play on an older card with little spare performance, or the game shows new problems with the newer model, stay with the older one. Some older games were tuned around the CNN look and can behave oddly when forced to a newer model.
A simple test
Play ten minutes with each model in a scene you know well. Watch fences, foliage, hair and fast turns. Keep the one you prefer.
A visual checklist for comparing the two models
You can judge the models yourself in about fifteen minutes. Use scenes that stress the weaknesses of each approach.
Scenes that reveal differences
Look for foliage swaying in the wind, wire fences and railings, hair and fur, distant buildings with regular windows, text on signs, and objects moving across detailed backgrounds. These scenes show shimmering and ghosting most clearly.
How to capture a fair comparison
Use the same position, the same time of day and the same mode. Take screenshots if the game allows, but also judge in motion, because many differences only show when the camera moves.
Questions to ask
Is the edge of a moving object cleaner? Do thin lines stay stable? Does text look sharper? Is anything worse? Write short notes, because impressions fade quickly.
How the models work together with other DLSS features
The model choice affects more than the upscaled picture.
Ray Reconstruction
Ray Reconstruction has its own model, and DLSS 4 introduced a transformer version for it. The two models are chosen separately in the app, so you can use a new upscaler with an older Ray Reconstruction or the reverse.
Frame generation
Frame generation uses a separate model too. NVIDIA's DLSS 4.5 update introduced a newer one for RTX 40 and RTX 50 cards. Read our guide to Multi Frame Generation for how those frames are made.
Why separate models matter
Each model can be updated independently, which is why a single game can show different versions in different parts of the DLSS family.
Why researchers moved from CNN to transformer designs
The shift in DLSS mirrors a wider shift across artificial intelligence.
The attention idea in plain terms
Attention lets a model decide which parts of its input to focus on. In language, that means linking a pronoun to a noun many words earlier. In an image sequence, it means linking a pixel in the current frame to the right pixel in a previous frame, even when the object moved far.
Why it suits games
Games contain fast motion, repeated patterns and sudden changes. A model that can look broadly and focus selectively handles these cases better than one limited to small windows.
The trade-off
Attention is more expensive to compute, which is why powerful Tensor cores matter. The fifth-generation Tensor cores in the RTX 50 series are built with this kind of work in mind, as the Blackwell architecture guide explains.
Four example owners and their choices
These examples are illustrations.
An RTX 3060 owner at 1080p
Frame rate is tight, and Quality mode is the safe choice. The owner tries the transformer model and finds a small frame rate drop but steadier foliage. They keep it in slower games and use the older model in one fast shooter where every frame counts.
An RTX 4070 owner at 1440p
The owner switches to the latest model in every DLSS game through the app, and moves from Quality to Balanced to recover a few frames, finding the picture still looks better than before.
An RTX 5080 owner at 4K
The owner uses the transformer model with Performance mode, because the model's gains are largest at low internal resolutions, and enjoys cleaner detail than the old Quality mode.
A laptop RTX 5060 owner
Heat is a concern, so the owner caps the frame rate, keeps the transformer model on and uses Balanced mode. The model's efficiency on recent hardware makes this comfortable.
What they share
Each owner tested rather than assumed, and each kept the freedom to switch back.
Troubleshooting model changes
| Symptom | Possible cause | Try this |
|---|---|---|
| Game crashes after forcing a model | Incompatible override | Set the app override back to the game's own setting |
| Picture looks worse | Game tuned for the older model | Compare modes, keep the better one |
| Frame rate drops a lot | Older card with a heavier model | Use the first-generation transformer or the CNN model |
| Override option missing | Old app or driver, or unsupported game | Update the app and driver |
| Indicator shows an old version | Override did not apply | Restart the game, confirm the setting |
Safe habits
Change one setting at a time, restart the game after each override change, and avoid overrides in online games with strict anti-cheat unless the game explicitly allows them.
How the model names map to preset letters
DLSS selects models through lettered presets. The details are covered in the dedicated guide to DLSS presets, and a short summary helps here.
CNN era presets
The CNN-based versions of DLSS used presets labelled with early letters such as A to F, each tuned for a different situation, and games usually picked one for you.
Transformer era presets
DLSS 4 introduced transformer presets, commonly seen as J and K, with K widely described as the better general choice. DLSS 4.5 added a second-generation transformer model, which independent coverage links to presets L and M. NVIDIA's own app is the authority on which letters your version offers.
Why the letters matter
If a guide tells you to "use preset K", it is naming a specific model. Without knowing the letter, "the transformer model" can mean more than one thing, which is why mixed advice online can confuse people.
The second-generation transformer in DLSS 4.5
NVIDIA announced DLSS 4.5 at CES 2026. Its Super Resolution model is described as a larger, second-generation transformer that runs on all RTX cards. Reports from NVIDIA and independent outlets say it improves sharpness and detail, especially in lower-resolution modes, at a modest performance cost on RTX 40 and 50 cards.
Who it suits
RTX 40 and RTX 50 owners get the best balance. Owners of RTX 20 and RTX 30 cards may see a bigger frame rate drop, so many testers recommend staying with the first-generation transformer unless a game looks clearly better and still runs well.
How to check what you have
The NVIDIA App's override page and the on-screen indicator show which model is active. Do not rely on a game's own label, because a game may name its options differently from NVIDIA.
Keep expectations realistic
A newer model is not a promise of a perfect picture. Games vary, and some will look better with one model than another. Treat the choice as something to test, not something to assume.
How the choice plays out by game type
Genres put different demands on the model.
Open-world adventures
These games have vast distances, foliage and fine detail. The transformer model's steadier edges are a clear benefit, and the performance cost is easy to absorb at moderate frame rates.
Racing games
Fast movement and repeated patterns, such as barriers and road markings, can shimmer with older models. The transformer model tends to keep these steadier, though ghosting around cars can still appear in some titles.
Competitive shooters
Clarity in motion matters, and so does the frame rate. Many players test both models and pick the one that gives clearer targets without losing frames.
Horror and dark scenes
Dark games with low light can show noise and smearing. Newer models can help, and Ray Reconstruction matters here too, as our guide to Ray Reconstruction explains.
Ray tracing and the transformer model
Games with ray tracing or path tracing benefit most from the combination of a transformer upscaler and a transformer Ray Reconstruction model. The upscaler rebuilds sharpness, while Ray Reconstruction cleans noisy lighting. Together they reduce the smeared look that early ray-traced games sometimes showed. The site's article on ray tracing versus path tracing explains why these games are so demanding.
A practical tip
If a ray-traced game looks smeary, check both the upscaler model and the Ray Reconstruction model in the app. They are separate settings and may be set differently.
Comfort and personal preference
People perceive upscaling artefacts differently. Some notice shimmering on thin lines immediately, while others never see it. A small number of players find flicker or ghosting uncomfortable over long sessions.
What to do if something bothers you
Try a higher quality mode, a newer model, or native resolution with DLAA. If you play with a controller on a large screen, small artefacts matter less than if you sit close to a desktop monitor.
Trust your own eyes
Reviews and charts guide you, but your own screen, distance and sensitivity decide what looks right. Spend a short time comparing, then stop and enjoy the game.
How to read comparison videos and articles critically
Plenty of online content compares DLSS models. Some of it is excellent, and some is misleading. A few habits help you tell the difference.
Check that the comparison is fair
A fair comparison uses the same game version, the same scene, the same internal resolution and the same driver. If the creator does not state these, you cannot trust that differences come from the model alone.
Watch for compression
Video platforms compress footage heavily, which can hide or exaggerate differences in fine detail. Zoomed side-by-side clips from an original capture are more reliable than a compressed video seen at normal size.
Look at motion, not only stills
A still frame can look identical while motion shows ghosting or shimmering. The best comparisons show the same camera movement for each model, slowed down so you can see the edges.
Notice the preset and the mode
A statement such as "the new model looks better" means little without the mode, because gains differ between Quality and Performance. Good creators state the resolution, the mode and the preset letter.
Remember your own conditions
Your screen size, viewing distance and eyesight differ from the reviewer's. Use their findings to decide what to test, then confirm on your own screen.
Support and driver notes
Keeping your software current avoids many model problems.
Drivers and the NVIDIA App
New models and overrides arrive through driver and app updates. Keep both reasonably current, and read the release notes for DLSS entries. The explainer on Game Ready and Studio drivers helps you pick which driver type to follow.
Game patches
Developers update the DLSS files bundled with their games at their own pace. If a game's patch notes mention a DLSS update, test it. If they do not, the NVIDIA App override can fill the gap, within the limits of the game and its anti-cheat.
Online games and overrides
Some online games with strict anti-cheat do not allow modified files. The NVIDIA App's override works at the driver level and is designed to be safe, but if a game's rules are strict, check them first. The separate guide to DLL swapping covers the risks of the manual route.
A decision flow for choosing a model
Use this short flow when you are unsure.
- Is your frame rate comfortable with the game's default model? If yes, you can stop.
- Does the picture show ghosting, shimmering or softness you dislike? If yes, try the latest transformer model.
- Did the frame rate fall too much? If yes, move to the next faster DLSS mode, or use the first-generation transformer.
- Is the picture still worse than the old model? If yes, go back to the default for that game.
- Write down what you chose, and re-check after major updates.
Why a flow beats a rule
No single rule works for every card, game and screen. A short test, repeated when something changes, gives you a better answer than any universal recommendation, including this one.
Frequently misunderstood points
A few misunderstandings come up in almost every discussion of DLSS models.
"The transformer model is a different feature"
It is not a new feature. It is a better engine inside the same feature. When you switch models, the settings menus look the same, and only the quality and the cost change.
"Newer is always sharper"
Newer models are usually cleaner and more stable, but sharpness is a matter of taste. Some people prefer a softer look, and some games add their own sharpening on top. Judge the picture, not the version number.
"Only the latest card benefits"
The transformer upscaler benefits every RTX card. The cards that benefit most are those with the spare performance to absorb its cost.
"I must update every game manually"
You do not. The NVIDIA App override can apply a newer model to many games without editing files, which is safer and quicker than manual changes.
Real games that show the difference
Several modern games ship with transformer models or accept them through the NVIDIA App. The site's list of top DLSS 4 and ray tracing games is a good place to pick a title for your own comparison, and the Fortnite comparison of DLSS 4 and FSR 3 shows how upscaler choices appear in a busy competitive game. Use them as starting points, then test on your own screen.
A short note on honesty
Everything above reflects NVIDIA's announcements and independent testing at the time of writing. Software features change, and details such as preset letters and recommended settings may shift. When this page and your NVIDIA App disagree, trust the app, and treat this page as an explanation of the ideas behind the settings.
What the transformer shift tells us about the future
DLSS is a fast-moving feature, and the move from CNN to transformer models is part of a long trend. NVIDIA has said it trains its models on powerful supercomputers and delivers improvements through software, which means a graphics card you buy today may look better in games next year. That is a reason to keep your software current. It is not a reason to buy a card for features that do not yet exist. The sensible approach is to buy for what a card does well today, enjoy improvements as they arrive, and ignore promises that depend on future game support.
How this connects to your buying decision
If you are choosing between cards, remember that the transformer upscaler works on every RTX card, so the image quality gain is not a reason to pay extra for the newest generation. The reasons to choose a newer card are performance, memory and features such as Multi Frame Generation. Our guide to the RTX 50 series can help you weigh those points.
When to re-check your model choice
Models improve, so a setting that was right last year may not be now.
- After a major driver or NVIDIA App update.
- After a game patch that mentions DLSS.
- When you change your monitor or resolution.
- When you move to a new graphics card.
A light-touch routine
Keep a note of the model and mode that worked for each game. When something changes, spend ten minutes testing instead of rebuilding your settings from scratch.
Short glossary for this article
- CNN: convolutional neural network, a model that looks at small neighbourhoods.
- Transformer: a model that weighs relationships across the whole input.
- Attention: the mechanism that lets a transformer focus on relevant parts.
- Preset: a lettered setting that selects a specific AI model.
- Ghosting: faint trails behind moving objects.
- Tensor core: the hardware unit that runs the AI model.
Frequently asked questions about DLSS models
What is the difference between the CNN and transformer DLSS models?
The CNN model looks at small neighbourhoods of pixels, while the transformer model weighs relationships across the whole picture. NVIDIA introduced transformer models with DLSS 4, and they generally give steadier edges, cleaner detail and less ghosting. They cost slightly more performance because the model is larger.
Does the transformer model work on RTX 30 series cards?
Yes. The transformer models for Super Resolution and Ray Reconstruction run on all RTX cards, including the RTX 20 and RTX 30 series. They may cost a little more performance on older cards, so check your frame rate. Frame generation features are separate and have their own hardware requirements.
Is the transformer model always better looking?
Usually, but not in every game. Most testers found improved stability and detail, and some older titles tuned around the CNN look behave oddly with a newer model. Compare both in your own games and keep whichever you prefer. The switch can be reversed in moments.
Will the transformer model lower my frame rate?
Slightly, in many cases. NVIDIA has described the cost on RTX 40 and 50 series cards as small, while older cards can lose more. If you notice a drop, try the next faster DLSS mode, which often returns the frame rate while keeping better quality than the old model.
How do I know which DLSS model a game uses?
Turn on NVIDIA's on-screen DLSS indicator, which can show the version and preset in use. Games that list DLSS 4 or later in their patch notes usually use the transformer model. If you used the NVIDIA App override, the indicator confirms whether the override applied.
Do I need an RTX 50 card for the transformer model?
No. Any RTX card can use the transformer models for upscaling and Ray Reconstruction. RTX 50 series cards are needed only for Multi Frame Generation. Owners of older cards therefore get the main image quality benefit of DLSS 4 without buying new hardware.