DLSS 4.5: one frame rendered, five generated

On March 31, NVIDIA shipped DLSS 4.5's 6X mode, and the name describes the output rather than the work. For every frame the card actually renders, a model produces five more. NVIDIA's phrasing: "generate five additional frames for every single natively rendered one, for a maximum 6X multiplier."¹

The full arithmetic is also NVIDIA's. Announcing DLSS 5 in March, the company summed up the previous generation by saying DLSS 4.5 "uses AI to draw 23 out of every 24 pixels seen on the screen."² Twenty-three out of twenty-four. That is what you get when upscaling that reconstructs three quarters of the pixels in each frame meets a multiplier in which five of every six frames never passed through the game engine.

None of that is an accusation; it is the product description, published by the vendor. What deserves a look is what the arithmetic delivers, what it charges, and what it removes from the bar chart.

Three releases wearing one number

DLSS 4.5 arrived in pieces across 2026, and mixing up the pieces is the usual mistake.

At CES on January 5 came the second-generation transformerA neural network architecture that weighs the relationship between all its inputs at once. In DLSS it reconstructs the image from samples taken across several frames. model for Super Resolution, the part that reconstructs the image. It works on every GeForce RTX card from the RTX 20 series on, and covered more than 400 games at launch.¹ ³

On March 31 came Dynamic Multi Frame GenerationA technique that manufactures in-between frames from two frames the game engine rendered, instead of computing each one. and 6X mode, exclusive to the RTX 50 series, requiring driver 595.97 or newer and switched on in the NVIDIA app under "DLSS Override – Frame Generation Mode."⁴

At Computex on May 31, the company announced the third piece: a second-generation transformer for Ray Reconstruction as well, with "35% more compute capability" and 20% more parameters at similar performance, promised for August across 27 games and again for every RTX card.⁵

These are different products under one label. The first improves the image on nearly any card in the line. The second multiplies the number on screen and exists only on the most expensive generation.

What the new model does to the pixels

The second-generation Super Resolution model uses "five times the compute power of the original transformer model" and was trained on an expanded dataset. The most concrete change is where it works: in linear space, which NVIDIA calls "the game engine's native ground truth," instead of the logarithmic space it used before. The payoff shows up in high-contrast scenes, where neon signs and bright reflections keep their full color range. The company says Performance mode is now comparable to native and that Ultra Performance has become viable at 4K.¹

There is fine print. The larger models, M and L, lean on FP8An 8-bit numeric format that speeds up running AI models. Cards without native support fall back to larger, slower formats. to run cheaply, and the RTX 20 and 30 series have no native support for the format: "Models M and L carry a heavier performance impact for these GPUs." For those cards NVIDIA recommends staying on Model K, the DLSS 4.0 one.¹ Owners of older hardware get access to the improvement and pay more for it.

Where rendering ends

Reconstruction and frame generation are operations of a different kind, and that distinction matters more than any quality comparison between them.

When Super Resolution turns 1080p into 4K, it starts from a frame the engine genuinely computed for that instant, with your mouse input already folded in, and uses history and motion vectorsData the game engine hands the GPU describing where each part of the scene moved between one frame and the next. to fill in what the lower-resolution render left out. It is sparse sampling reconstructed, the same family of trick temporal anti-aliasing has been doing for years, with a far better model where the heuristic used to sit.

Frame generation does something else. It produces an image for an instant the engine never simulated. Nobody sampled your mouse at that moment, nobody stepped the physics, nobody asked the server for state. The frame is a plausible interpolation between two real states, and plausibility is the whole guarantee.

That is why the two fail in opposite ways. Bad reconstruction shows up as blur or aliasing, and you keep playing. Bad generation shows up as a response that doesn't match what the screen is showing you.

Latency doesn't follow the multiplier

The strongest argument against reading the big number sits on NVIDIA's own slide. In Black Myth: Wukong with path tracingThe most complete form of ray tracing: it simulates light travelling through the whole scene rather than applying precomputed lighting., 4X mode delivers 184 FPS off 46 frames per second actually rendered, at 47 ms of latency. 6X mode delivers 246 FPS at 53 ms.⁶

Divide 246 by six and you get 41 rendered frames per second. Moving from 4X to 6X, the counter gained 62 FPS, the card lost five real frames per second because generating them costs GPU time of its own, and the response got 6 ms worse. The chart the vendor shows is the 246 one.

Independent testing reaches the same place from another direction. HotHardware measured an RTX 5070 Ti on a 4K 160 Hz display using NVIDIA's own Frameview and found latency essentially flat across multipliers; in Nioh 3, with Reflex on, "nearly no increase in latency" even as Dynamic mode pushed the average to 164.7 FPS.⁷

The explanation is simple: the multiplier changes how many images appear, not when the game reads your mouse. Latency is set by the rendered frame rate, and that rate doesn't rise. At 6X it drops slightly. That is why HWCooling's technical write-up noted latency getting worse precisely as the on-screen number takes off.⁶

What still breaks on screen

The artifacts that remain follow a clear pattern: they live wherever motion vectors don't.

In HotHardware's testing, Cyberpunk 2077's flat display screens, animated 2D textures with no geometry behind them, came out "smeared into illegibility" in places; in Hogwarts Legacy, the magical paintings are MP4 videos playing on a wall, and the interpolator, given no displacement data at all, smears the whole thing. In Nioh 3, in the same session, the reviewer saw no artifacts whatsoever.⁷ The difference isn't luck. It comes down to whether the engine hands the model any information about where that content moved.

There is also a floor. The technique needs a decent base rate, somewhere around 40 FPS or more, and the same test is blunt about what happens below it: "when the framerate dips with FG on, you really feel it."⁷ Multiplying 25 FPS by six does not produce a 150 FPS game; it produces a 25 FPS game with 150 images.

Dynamic Multi Frame Generation goes straight at that discomfort. Rather than fixing the multiplier, it tracks the gap between what the GPU is delivering and the display's refresh rate, and generates only what's needed. NVIDIA compares it to an automatic transmission.¹ ⁴ It is the least showy part of the release and probably the most useful: in independent testing the automatic mode came out slightly more responsive than fixed 4X, and the reviewer treated it as something to switch on and forget.⁷

How benchmarks should report this

What breaks the comparison isn't DLSS existing, it's two numbers with different meanings sharing one cell in the table.

The best-documented case is May 2025, at the RTX 5060 launch. NVIDIA made early access conditional on comparisons against the RTX 3060 and RTX 2060 Super, at 1080p, with Multi Frame Generation 4X enabled, against cards that don't have the technology. Hardware Unboxed described the result as a synthetic advantage that doesn't correspond to raw performance, and Gamers Nexus, given no review sample, bought a card to test outside those conditions.⁸ Earlier, the claim that a $549 RTX 5070 would deliver RTX 4090 performance only held with the multiplier running on one side and not the other; in a game without support, the 4090 pulls comfortably ahead.⁹

The honest format isn't hard, and NVIDIA's own slide already contains it: rendered rate, displayed rate and measured latency, always together. Publish only the third column and you are reporting the multiplier, not the card.

What the community picked in the dark

The reconstruction half won its argument comfortably. ComputerBase ran a blind test across six games at 4K (Anno 117, ARC Raiders, Cyberpunk 2077, Horizon Forbidden West, Satisfactory and The Last of Us Part II), comparing DLSS 4.5 and FSR Upscaling AI on the Quality preset against native resolution with each game's own TAA, with the clips labelled only 1, 2 and 3 and shuffled between rounds.¹⁰ Of 6,747 votes, 3,249 of them, or 48.2%, went to DLSS 4.5, which won in all six games. Native with TAA and FSR trailed.¹¹

In the comments, readers raised the sharpest objection to the test itself: TAA is a weak baseline, and part of what upscaling wins comes from native being saddled with poor anti-aliasing.¹⁰ It is a fair caveat, and it doesn't undo the result. Reconstructing well is now the house standard; the other half of the package is what remains contested.

The price of being allowed to use it

Here the subject leaves the chart and lands in your wallet. 6X and Dynamic mode exist only on RTX 50 cards,¹ ⁴ and in August 2026 those cards cost considerably more than they did in June. PCGamesN's survey of median Newegg prices puts the 16 GB RTX 5060 Ti at $804.99, up from $569.99, and the RTX 5070 at $899.99, up from $659.99, over two months, with the AI-driven memory shortage named as the cause.¹²

This is the same memory crunch that has been keeping the Super refresh in a drawer. The practical result is an awkward inversion: the feature sold as free performance only switches on inside hardware whose street price rose 36% to 40% in two months. For anyone outside the current generation, DLSS 4.5 is the image upgrade, which is no small thing and costs nothing.

Verdict

The reconstruction model is the best part of this release, and it's the part almost everyone gets. It improves the image without demanding a new card, and a blind test with thousands of participants preferred its output to native rendering in all six games.¹¹

6X mode is narrower than the marketing suggests. It serves one case well: a comfortable base frame rate, a 240 Hz or faster display, a single-player game, and a gap to fill. Outside that case, the multiplier delivers a higher count without the responsiveness that count would normally imply, and NVIDIA publishes both numbers side by side while promoting only one.

What stays unresolved is the vocabulary. As long as "FPS" covers both what the card computed and what the model drew, comparisons across generations and across vendors lose their meaning. And the pressure is only building: DLSS 5, announced on March 16 and promised for this fall, puts AI into lighting and materials too, and Jensen Huang himself called it "the GPT moment for graphics."² The more of the final image is inferred, the more "how many frames per second?" will need a second question attached: how many of them did your card actually compute?

Sources

1. NVIDIA DLSS 4.5 Delivers Major Upgrade With 2nd Gen Transformer Model For Super Resolution & 6X Dynamic Multi Frame Generation (second-generation transformer using "five times the compute power of the original transformer model" and an expanded dataset; operating in linear space, "the game engine's native ground truth"; Performance mode comparable to native and Ultra Performance viable at 4K; Models K, M and L, with "Models M and L carry a heavier performance impact for these GPUs" on the RTX 20 and 30 series for lack of native FP8; "generate five additional frames for every single natively rendered one, for a maximum 6X multiplier"; up to 35% higher 4K frame rates in path-traced titles moving from 4X to 6X; Dynamic MFG compared to an automatic transmission; Super Resolution for all RTX cards across 400+ games, Dynamic MFG and 6X for RTX 50 only) · NVIDIA · https://www.nvidia.com/en-us/geforce/news/dlss-4-5-dynamic-multi-frame-gen-6x-2nd-gen-transformer-super-res/ · 2026-01.
2. NVIDIA DLSS 5 Delivers AI-Powered Breakthrough in Visual Fidelity for Games (NVIDIA describing DLSS 4.5 as technology that "uses AI to draw 23 out of every 24 pixels seen on the screen"; DLSS 5 taking each frame's color and motion vectors and using an AI model to infuse photoreal lighting and materials "anchored to source 3D content"; "DLSS 5 will arrive this fall"; Jensen Huang calling it "the GPT moment for graphics") · NVIDIA Newsroom · https://nvidianews.nvidia.com/news/nvidia-dlss-5-delivers-ai-powered-breakthrough-in-visual-fidelity-for-games · 2026-03-16.

Show 10 more sourcesHide sources

3. GeForce @ CES 2026: NVIDIA DLSS 4.5 Announced (DLSS 4.5 announced with a "2nd generation transformer model for DLSS Super Resolution"; Super Resolution available immediately to all GeForce RTX owners; 6X Dynamic Multi Frame Generation promised for spring, restricted to RTX 50 Series) · NVIDIA · https://www.nvidia.com/en-us/geforce/news/ces-2026-nvidia-geforce-rtx-announcements/ · 2026-01-05.
4. DLSS 4.5 Dynamic Multi Frame Generation and 6X Mode Released (released 2026-03-31 through an NVIDIA app update, requiring GeForce Game Ready Driver 595.97 WHQL or newer; enabled under "DLSS Override – Frame Generation Mode" with a choice of "Max refresh rate" or a custom target; the system continuously monitors the gap between GPU performance and the display's maximum refresh rate; 6X restricted to RTX 50 Series) · NVIDIA · https://www.nvidia.com/en-us/geforce/news/dlss-4-5-dynamic-multi-frame-generation-6x-mode-released · 2026-03-31.
5. DLSS 4.5 Ray Reconstruction Announced; Over 1000 RTX Games & Apps Available Now (second-generation transformer for Ray Reconstruction with "35% more compute capability, and processes 20% more parameters, while maintaining similar performance"; scheduled for August 2026 across all GeForce RTX GPUs; 27 games listed) · NVIDIA · https://www.nvidia.com/en-us/geforce/news/dlss-4-5-ray-reconstruction-1000-rtx-games-apps-out-now/ · 2026-05-31.
6. DLSS 4.5 brings 6× frame generation and better image quality (figures from NVIDIA's presentation for Black Myth: Wukong, quoted as "4× generation results in 184 FPS (46 FPS in-game) with a latency of 47 ms, while 6× generation delivers 246 FPS with a latency of 53 ms"; analysis that latency worsens because it tracks the genuinely rendered rate rather than the displayed one; second-generation model trained on a larger dataset and slower on RTX 20 and 30 cards) · HWCooling · https://www.hwcooling.net/en/dlss-4-5-brings-6x-frame-generation-and-better-image-quality/ · 2026-01-06, updated 2026-01-24.
7. NVIDIA DLSS 4.5 Tested: The Real-World Impact Of Dynamic Multi-Frame Generation (independent testing on a GeForce RTX 5070 Ti with a 4K 160 Hz display, latency measured with Frameview; Hogwarts Legacy at a 65.7 FPS base with smearing on in-game MP4 playback; Cyberpunk 2077 display screens "smeared into illegibility"; Nioh 3 at 164.7 FPS in Dynamic mode with "nearly no increase in latency" and no artifacts observed; Dynamic mode slightly more responsive than fixed 4X; the caveat that "when the framerate dips with FG on, you really feel it" and a recommended base of roughly 40 FPS and up) · HotHardware · https://hothardware.com/news/nvidia-dlss-45-dynamic-mfg-tested · 2026-03.
8. Nvidia's RTX 5060 launch and the erosion of independent GPU reviews (conditions imposed on early access: no comparison against the RTX 4060, comparisons limited to the RTX 3060 and RTX 2060 Super, mandatory 1080p testing and mandatory 4x multi-frame generation; Hardware Unboxed's objection that this creates a synthetic advantage that doesn't translate to raw performance; Gamers Nexus given no review sample) · TechSpot · https://www.techspot.com/news/107962-nvidia-rtx-5060-launch-erosion-independent-gpu-reviews.html · 2025-05-19.
9. I just tested the Nvidia GeForce RTX 5070, and yes it can beat the RTX 4090, but there's a big catch (the favorable RTX 5070 comparison depends on multi frame generation on one side and ordinary frame generation on the other; "if a game doesn't support multi frame gen, then the RTX 4090 will be significantly faster than the RTX 5070, as the underlying GPU has so much more raw horsepower") · PCGamesN · https://www.pcgamesn.com/nvidia/geforce-rtx-5070-vs-4090 · updated 2025-03-05.
10. Nativ vs. DLSS 4.5 vs. FSR Upscaling AI im Leser-Blindtest (blind-test methodology: six games (Anno 117, ARC Raiders, Cyberpunk 2077, Horizon Forbidden West, Satisfactory and The Last of Us Part II) at Ultra HD, Quality preset for both upscalers, "Native" being native resolution plus each game's own TAA; external capture in AV1 at 60 Mbps played back in NVIDIA ICAT; clips labelled "1", "2" and "3" and shuffled; reader discussion of TAA as a weak baseline) · ComputerBase, Wolfgang Andermahr · https://www.computerbase.de/artikel/grafikkarten/nativ-vs-dlss-4-5-vs-fsr-upscaling-ai-leser-blindtest.95965/ · 2026-02-03, updated 2026-02-18.
11. Wochenrück- und Ausblick: Die Community katapultiert DLSS im Blindtest an die Spitze (blind-test result: "Insgesamt 6.747 Stimmen sind von der Community abgegeben worden, 3.249 Stimmen respektive 48,2 Prozent der Stimmen entfielen auf Nvidia DLSS 4.5"; DLSS 4.5 winning all six games) · ComputerBase · https://www.computerbase.de/news/in-eigener-sache/wochenrueckblick-und-ausblick-2026-w08.96242/ · 2026-02-22.
12. Nvidia RTX 50 graphics cards see drastic rises above retail price, as AI memory boom continues (median Newegg US prices between June and August 2026: 16 GB RTX 5060 Ti from $569.99 to $804.99 and RTX 5070 from $659.99 to $899.99; the AI-driven memory shortage named as the cause; top-end models up only single-digit percentages) · PCGamesN · https://www.pcgamesn.com/nvidia/rtx-50-graphics-cards-prices-august-2026 · 2026-08-12.