Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

AI Upscaling vs Preservation: Two Very Different Ways to Make an Old Game Look Better

2026-08-29 · Explainer · Updated 2026-08-29

There are two philosophies for making a 20 year old game look good on a 2026 display, and people talk about them as if they are the same activity. They are not.

One tries to reproduce the picture the game was designed to produce. The other tries to invent detail that was never there.

Both are legitimate. Both have real fans. But they answer different questions, they fail in different ways, and knowing which one you are doing saves a lot of arguing. The recent DLSS 5 leak, where modders put neural rendering into Skyrim and GTA San Andreas, has made the split unusually easy to see.

The Problem Both Approaches Are Solving

An old game looks wrong on a modern screen, and it is worth being precise about why.

A SNES output 256x224 pixels over an analog signal to a CRT. That CRT did things a modern panel does not. Scanlines with gaps between them. A phosphor mask giving every pixel visible internal structure. Bright areas blooming into dark ones. Horizontal blending that artists used deliberately to create colors and transparency effects the console could not produce directly.

Put that same 256x224 image on a 4K OLED and you get hard nearest neighbour squares, or a soft blurry mess, depending on how it was scaled. Neither is what anyone saw in 1994.

So something has to be done. The question is what.

Approach One: Reproduce the Intended Output

This is the preservation position. The game already looked correct. Your job is to recreate the conditions under which it looked correct.

The toolkit is mature:

Integer scaling. Multiply the source resolution by a whole number so every original pixel becomes an identical square block. No uneven pixels, no shimmer. The floor of doing this properly.

CRT shaders. Software recreations of the display, covering scanlines, phosphor masks and geometry. Our RetroArch shaders guide covers getting a good result for free.

Energy conservation scanlines. A refinement worth knowing about. Naive scanline shaders draw black lines over the picture, which dims everything. Better ones redistribute the light instead of removing it, keeping total brightness constant.

Phosphor and aperture grille masks. Simulating the physical structure of the tube, which is what gives real CRT output its texture.

HDR CRT simulation. The newest development and a genuinely big one. CRT phosphors were much brighter than the pixels around them, which on an SDR panel means choosing between a dim image and blown highlights. HDR gives you the headroom for both. This is a headline feature of current hardware scalers like the RetroTINK-6X CE.

Black frame insertion. CRTs had almost no motion blur, because the phosphor decayed between refreshes. Inserting black frames approximates that on a sample and hold LCD or OLED, trading brightness for motion clarity.

The unifying idea: nothing is invented. Every one of these takes information the game produced and presents it the way the hardware would have. If a sprite was 16 pixels wide, it is still 16 pixels wide.

Approach Two: Infer New Detail

The other position says the original picture was a limitation, not a target. Modern hardware can do better, so it should.

AI texture upscaling. A model takes a 64x64 texture and generates a 512x512 version, inventing detail the artist never drew. Popular in Skyrim and Morrowind modding for years.

Spatial upscalers. DLSS, FSR and XeSS reconstructing a higher resolution image from a lower resolution one, using motion vectors and temporal data.

Neural rendering. The newest and most aggressive. DLSS 5 does not upscale the image, it re-infers how surfaces should look. Lighting, materials, subsurface scattering through skin, the way light catches fabric. NVIDIA's own description is that it infuses the scene with photoreal lighting and materials anchored to the source content.

The unifying idea: information is added. The output contains detail that does not exist in the source.

That is the whole difference, and it is not a small one.

Where It Goes Wrong

The DLSS 5 mod results are an unusually clean demonstration, so they are worth using.

Applied to Skyrim, it makes characters look photorealistic. It also makes them look like different people. Smoother skin, more conventional features, more even lighting. Critics landed on "beauty filter" and it is an accurate description.

Applied to GTA San Andreas, it reworks the lighting on a game whose hazy, sun bleached, deliberately low fidelity look was the point.

Here is the thing to understand: that is not a bug. The model is doing exactly what it was trained to do. It has one direction of travel, toward photorealism, and it does not know that Skyrim's characters look weathered on purpose, or that San Andreas is not trying to look like a 2026 game.

An AI model that adds detail must decide what detail to add. That decision is an artistic one. It is being made by a model trained on photographs, not by the people who made your game.

This is why the preservation crowd gets irritated by the framing. Calling a CRT shader and a neural filter both "making the game look better" hides the fact that one of them is making an aesthetic choice on the artist's behalf.

We covered the specific results, game by game, in every game the DLSS 5 mod has been tested on.

The Honest Middle Ground

There is a third thing, and most people reading this already use it.

Emulator internal resolution upscaling. When you set DuckStation or PCSX2 to render at 4x native, you are not inventing anything. The original 3D geometry is being rasterised at a higher resolution. Same polygons, same textures, same everything, drawn with more pixels.

This is genuinely different from AI upscaling, and the distinction matters. The game's 3D data always contained more precision than a PS1 could display. Rendering at higher resolution reveals information that was already there.

It is also not free of aesthetic consequence, which is worth being honest about. PS1 games used low resolution deliberately to hide rough geometry, and dithering patterns designed to blend on a CRT become visible crosshatching at 4x. Some games look worse upscaled. That is why serious emulation setups often pair high internal resolution with a CRT shader, using both approaches at once.

Which Should You Use?

Depends entirely on the game.

Use preservation tools for 2D games. Sprite art from the 8 and 16 bit eras was drawn pixel by pixel for a specific display. There is no hidden detail to recover, and everything an AI adds is invention. Integer scaling plus a good CRT shader is correct here and it is not close.

Use internal resolution upscaling for 3D games from the PS1, N64, Dreamcast and PS2 eras. The geometry supports it. Add a shader if dithering artifacts bother you.

Consider AI texture upscaling for older PC games if you are the sort of person who enjoys modding and does not mind the result drifting from the original. This is where it is most defensible, because the community has decades of practice and you can pick and choose.

Be cautious with neural rendering on anything stylised. As of now this is barely available anyway, but the principle holds. The more distinctive a game's look, the more a photorealism model will damage it.

And run whatever you actually enjoy. These are your games. Someone else's purity standard is not a rule you have to follow. The point of understanding the difference is so you are choosing, not so you are choosing correctly.

What About Handhelds?

Practically speaking, the preservation toolkit is the one available to you.

Every retro handheld can run RetroArch shaders, though weaker devices will want the lighter presets. Integer scaling is standard. Emulator internal resolution upscaling works on anything with the power for it, which for PS1 and Dreamcast is most modern devices.

Neural rendering is not on the table. DLSS is NVIDIA only, DLSS 5 has so far only been shown on RTX 50 series cards, and no NVIDIA GPU handheld exists. We go through that in detail in the DLSS 5 requirements piece.

Which is fine, honestly. For the games most people play on handhelds, the preservation approach is the correct one anyway.

Frequently Asked Questions

What is the difference between AI upscaling and CRT shaders?

CRT shaders reproduce how a game looked on the display it was designed for, using only the information the game produces. AI upscaling generates new detail that was never in the source. One is recreating an original picture, the other is inventing a new one.

Is emulator internal resolution upscaling the same as AI upscaling?

No. Internal resolution upscaling renders the game's original 3D geometry at a higher resolution, so nothing is invented. The precision was always in the game data, the original console just could not display it. AI upscaling generates detail that does not exist in the source at all.

Do CRT shaders make games look worse?

They make games look softer and dimmer than raw pixels, which some people dislike at first. That softness is usually the point, since artists used CRT blending deliberately to create colors and effects the hardware could not produce directly. Energy conservation scanlines and HDR CRT simulation address the brightness problem specifically.

Should I use AI upscaled textures in old games?

It is most defensible in PC game modding, where communities have decades of practice and you can pick individual texture packs. Be aware that the result drifts from the original art, and that AI generated detail is a creative choice being made by a model rather than by the game's artists.

Can retro handhelds do AI upscaling?

Not neural rendering. DLSS is NVIDIA only and no NVIDIA GPU handheld exists. Handhelds can run CRT shaders, integer scaling and emulator internal resolution upscaling, which covers the approaches that actually suit retro games.

Why did DLSS 5 in Skyrim get criticised so heavily?

Because it changed the characters rather than sharpening them. The model pushes everything toward photorealism, so Skyrim's deliberately weathered faces came out smooth and conventionally attractive, looking like different people. The model is not wrong, it just has a target the original artists did not choose.

What is the best way to make retro games look good on a modern TV?

For 2D games, integer scaling plus a good CRT shader. For 3D games from the PS1 era onward, raise the emulator's internal resolution and add a shader if dithering artifacts appear. If you are playing on original hardware rather than emulating, a hardware scaler handles it.