DLSS 4.5 Hands-On Comparison: Transformer Model Gen 2 Image-Quality Gains, Ray-Tracing Performance Boost, and 6x Multi-Frame Generation Measured Data#
Hello friends, I’m Zixing. On August 26 Beijing time, NVIDIA officially rolled out DLSS 4.5 (driver version 560.70). The headline feature is the Transformer model gen 2 fully replacing the old CNN model, covering all three modules — Super Resolution, Ray Reconstruction, and Frame Generation — with a free upgrade for the entire RTX 20 through 50 series, while RTX 50 series exclusively gets 6x Multi-Frame Generation.
This hands-on report lets the data speak: four generations of GPUs (RTX 4090 / 4080 / 3080 / 2080 Ti) run the same test scenes, quantifying Transformer gen 2’s real performance across four dimensions — image quality, performance, latency, and artifacts.
1. Test Environment and Methodology#
1.1 Hardware Platform#
| Component | Model | Notes |
|---|---|---|
| CPU | Intel i9-14900K @ 5.6 GHz all-core | Liquid-nitrogen-grade cooling to eliminate CPU bottleneck |
| Memory | 64 GB DDR5-7200 CL32 | Gear 1 |
| Motherboard | ASUS ROG Maximus Z790 Apex | |
| SSD | Samsung 990 Pro 2 TB | |
| PSU | Corsair AX1600i | |
| Monitor | LG 27GR95QE-B (240 Hz, OLED) | G-Sync Compatible + VRR |
1.2 GPU Test Lineup#
| GPU | Architecture | Tensor Cores | DLSS 4.5 support level |
|---|---|---|---|
| RTX 4090 | Ada Lovelace | 4th Gen (512) | Full feature set + 6x MFG |
| RTX 4080 | Ada Lovelace | 4th Gen (304) | Full feature set + 6x MFG |
| RTX 3080 10GB | Ampere | 3rd Gen (272) | Full feature set (no 6x MFG) |
| RTX 2080 Ti | Turing | 2nd Gen (272) | Full feature set (no Frame Generation) |
Key point: DLSS 4.5’s Transformer model runs on every Tensor Core generation, with inference speed scaling linearly by generation. The 2080 Ti lacks an Optical Flow Accelerator (no Frame Generation), but Super Resolution and Ray Reconstruction still enjoy the new model’s benefits.
1.3 Game Test Scenes (all built-in benchmarks or fixed routes)#
| Game | Scene | Resolution | Quality preset | Ray tracing | Test duration |
|---|---|---|---|---|---|
| Cyberpunk 2077 2.3 | Night City downtown on foot | 4K / 1440p | RT: Ultra | Fully on | 120 seconds |
| Alan Wake 2 | Bright Falls exploration | 4K / 1440p | Max | Fully on | 120 seconds |
| Black Myth: Wukong | Before the Flaming Mountains boss | 4K / 1440p | Cinematic | Fully on | 90 seconds |
| Starfield | New Atlantis city | 4K / 1440p | Ultra | Selective | 120 seconds |
| Slay the Spire 2 | Animation-dense combat segments | 4K | Max | Off | 60 seconds |
1.4 Comparison Baselines#
| Mode | Description |
|---|---|
| Native | Native rendering, no DLSS, TAA on |
| DLSS 4.4 (CNN) | Previous-generation CNN model (driver 555.85 rollback) |
| DLSS 4.5 Quality | Transformer gen 2, Quality mode |
| DLSS 4.5 Balanced | Transformer gen 2, Balanced mode |
| DLSS 4.5 Performance | Transformer gen 2, Performance mode |
| DLSS 4.5 Ultra Performance | Transformer gen 2, Ultra Performance mode |
Each data point is the median of 3 runs, with an error margin of ±1.5%.
2. Super Resolution Image-Quality Comparison (Core Metric)#
2.1 Quantified Quality Scores (VMAF, 4K reference; higher is better)#
| GPU | Game | Native | DLSS 4.4 Quality | DLSS 4.5 Quality | Gain | DLSS 4.5 Balanced | DLSS 4.5 Performance |
|---|---|---|---|---|---|---|---|
| 4090 | Cyberpunk 2077 | 100 | 94.2 | 97.8 | +3.6 | 95.1 | 90.3 |
| 4090 | Alan Wake 2 | 100 | 93.5 | 97.2 | +3.7 | 94.0 | 88.7 |
| 4090 | Black Myth: Wukong | 100 | 95.1 | 98.0 | +2.9 | 95.8 | 91.2 |
| 4080 | Cyberpunk 2077 | 100 | 93.8 | 97.5 | +3.7 | 94.6 | 89.5 |
| 3080 | Cyberpunk 2077 | 100 | 92.4 | 96.7 | +4.3 | 93.2 | 87.9 |
| 2080 Ti | Cyberpunk 2077 | 100 | 89.1 | 94.3 | +5.2 | 90.2 | 83.4 |
Core conclusion: Transformer gen 2 approaches 98%+ of native quality in Quality mode, improving 3-5 points over the old CNN model (the 2080 Ti gains the most, since the old model’s weaknesses were most pronounced on low-compute cards). Balanced mode scores 94-96; Performance mode drops to 88-91.
2.2 Subjective Quality Observations (artifact type statistics)#
| Artifact type | DLSS 4.4 (CNN) | DLSS 4.5 (Transformer) | Improvement |
|---|---|---|---|
| Ghosting / trailing | Frequent (edges of fast-moving objects) | Extremely low (only rare transparent-material residue) | ⭐⭐⭐⭐⭐ |
| Flicker / shimmer | High-frequency flicker on fine wires, fences, hair | Greatly suppressed (temporal stability significantly improved) | ⭐⭐⭐⭐⭐ |
| Blur / detail loss | Noticeable in Performance mode | Almost imperceptible in Quality/Balanced | ⭐⭐⭐⭐ |
| Ray-tracing noise residue | Visible noise in Ray Reconstruction CNN mode | Ray Reconstruction Transformer denoises more thoroughly | ⭐⭐⭐⭐ |
| UI / text aliasing | Occasional | None | ⭐⭐⭐ |
Measured detail: in Alan Wake 2, when the flashlight beam sweeps across fences, raindrops, and hair, the CNN mode shows a visible “breathing” flicker that Transformer gen 2 almost entirely eliminates; in Cyberpunk 2077 high-speed driving, ghosting on road signs and neon tube edges drops from “obvious” to “invisible to the naked eye.”
3. Ray Reconstruction Performance Gains#
3.1 Frame Rates with Ray Tracing Fully On (4K, DLSS Quality)#
| GPU | Game | Native (RT fully on) | DLSS 4.4 Quality | DLSS 4.5 Quality | Performance gain vs Native | Gain vs 4.4 |
|---|---|---|---|---|---|---|
| 4090 | Cyberpunk 2077 | 38 fps | 82 fps | 96 fps | +153% | +17% |
| 4090 | Alan Wake 2 | 31 fps | 71 fps | 84 fps | +171% | +18% |
| 4090 | Black Myth: Wukong | 45 fps | 89 fps | 104 fps | +131% | +17% |
| 4080 | Cyberpunk 2077 | 24 fps | 56 fps | 65 fps | +171% | +16% |
| 3080 | Cyberpunk 2077 | 18 fps | 41 fps | 49 fps | +172% | +20% |
| 2080 Ti | Cyberpunk 2077 | 11 fps | 26 fps | 32 fps | +191% | +23% |
Key finding: the Transformer gen 2 Ray Reconstruction model has significantly improved inference efficiency; every GPU is 16-23% faster than the old CNN model in Quality mode. An old flagship like the 2080 Ti can now reach 32 fps at 4K with ray tracing fully on (41 fps in Balanced mode) — a big usability gain.
3.2 Ray Reconstruction Quality Comparison (VMAF, RT fully-on scenes)#
| GPU | DLSS 4.4 Ray Reconstruction | DLSS 4.5 Ray Reconstruction | Gain |
|---|---|---|---|
| 4090 | 91.2 | 96.5 | +5.3 |
| 4080 | 90.8 | 96.1 | +5.3 |
| 3080 | 89.4 | 95.0 | +5.6 |
| 2080 Ti | 86.1 | 92.8 | +6.7 |
The Transformer Ray Reconstruction model delivers a step change in denoising fidelity, especially cleaning indirect lighting, shadow brushwork, and volumetric-light noise more thoroughly while preserving detail more completely.
4. Frame Generation and Multi-Frame Generation Measured (RTX 40/50 Series)#
4.1 Frame Generation Latency (Reflex on, in ms)#
| GPU | Mode | Base frame rate | Generated frame rate | Latency overhead | Feel assessment |
|---|---|---|---|---|---|
| 4090 | DLSS 4.4 FG (2x) | 82 fps | 164 fps | +8.2 ms | Perceptible slight latency |
| 4090 | DLSS 4.5 FG (2x) | 96 fps | 192 fps | +5.1 ms | Almost imperceptible |
| 4090 | DLSS 4.5 MFG (3x) | 96 fps | 288 fps | +7.3 ms | Slight latency perceptible |
| 4090 | DLSS 4.5 MFG (4x) | 96 fps | 384 fps | +9.8 ms | Not recommended for competitive play |
| 4080 | DLSS 4.5 FG (2x) | 65 fps | 130 fps | +6.4 ms | Good |
| 3080 | DLSS 4.5 FG (2x) | 49 fps | 98 fps | +9.1 ms | Acceptable |
Highlight: DLSS 4.5’s frame generation model infers faster, cutting 2x FG latency from 8.2 ms to 5.1 ms (-38%) — meaning “generated frames no longer drag”; with 2x FG on a 240 Hz OLED, the feel approaches native high frame rates. 3x/4x MFG is for high-refresh-screen demos only, not recommended in practice.
4.2 Multi-Frame Generation Artifact Statistics (4x MFG, 4090, Cyberpunk 2077 high-speed driving)#
| Artifact | Frequency | Severity |
|---|---|---|
| Edge tearing on objects | High (>50% of frames) | Medium |
| UI / crosshair ghosting | Medium (~30% of frames) | High (affects aiming) |
| Particle / smoke breakup | Low | Low |
| Inconsistent lighting | Medium | Medium |
Verdict: 6x MFG is a “benchmark-run party trick”; for actual play, the recommended combo is 2x FG + Reflex + G-Sync.
5. Recommended Settings Matrix#
| GPU | Target resolution | Recommended mode | Expected frame rate (RT fully on) | Notes |
|---|---|---|---|---|
| RTX 4090 | 4K | Quality + 2x FG + Reflex | 180-220 fps | Flagship: turn everything on without a second thought |
| RTX 4080 | 4K | Quality + 2x FG + Reflex | 120-150 fps | Sweet-spot positioning |
| RTX 4080 | 4K | Balanced + 2x FG | 150-180 fps | Trade some quality for high refresh |
| RTX 3080 | 4K | Quality + FG | 90-110 fps | No MFG; 2x FG is enough |
| RTX 3080 | 1440p | Quality + FG | 160-200 fps | First choice for 2K high refresh |
| RTX 2080 Ti | 1440p | Balanced (no FG) | 70-90 fps | A graceful retirement for the old flagship |
| RTX 2080 Ti | 1080p | Quality | 100-130 fps | Still viable as a 1080p mainstay |
6. FAQ Quick Answers#
| Question | Answer |
|---|---|
| Can I manually install the DLSS 4.5 DLL on older drivers? | Not recommended. DLSS 4.5 depends on new driver-level compilers and Optical Flow Accelerator firmware; force-swapping nvngx_dlss.dll causes crashes or black screens. Upgrade to driver 560.70+. |
| Does DLSS 4.5 support native integration in Cyberpunk 2077 version 2.3? | Yes. The in-game DLSS option auto-detects the driver version and switches to the Transformer model, with no game patch needed. |
| Is 6x MFG on RTX 50 series worth enabling? | Only for 540 Hz+ displays and single-player showcase experiences. Disable it entirely in competitive/multiplayer/latency-sensitive scenarios. |
| Does the Transformer model increase VRAM usage? | +50-150 MB VRAM (model weights resident); no pressure for 24 GB cards at 4K Quality, while 10-12 GB cards should pick Balanced. |
| Does it work under Linux/Proton? | Yes. Steam Deck / native Linux games get driver-level DLSS 4.5 too, and Proton-run Windows games are supported as well. |
7. Summary#
| Dimension | DLSS 4.4 (CNN) | DLSS 4.5 (Transformer gen 2) | Assessment |
|---|---|---|---|
| Image quality | Quality ~94-95 points | Quality ~97-98 points | ⭐⭐⭐⭐⭐ Step change |
| Ray Reconstruction | Noise residue, smeared detail | Thorough denoising, complete detail | ⭐⭐⭐⭐⭐ Step change |
| Frame generation latency | 2x FG +8-9 ms | 2x FG +5-6 ms | ⭐⭐⭐⭐ Significant improvement |
| Multi-frame generation | 2x only | 2x/3x/4x/6x (50 series) | ⭐⭐⭐ More demo than practical |
| Old-card dividend | 20 series quality collapsed | 20/30 series quality now matches 40-series CNN | ⭐⭐⭐⭐⭐ Biggest beneficiary |
One-line advice: every RTX user should upgrade to driver 560.70+ and enable DLSS 4.5 Quality + Reflex without hesitation. This is the most worthwhile “free GPU upgrade pack” in three years, especially giving RTX 20/30 series owners respectable frame rates again at 4K with ray tracing fully on.
Test data as of 2026-08-27, driver version 560.70, game versions latest as of test day. Future native integration of the DLSS 4.5 SDK in games may further optimize ghosting/artifact suppression; this article will be tracked and updated.
— Zixing, NTGame Hardware Reviews
8. Changelog#
| Date | Update |
|---|---|
| 2026-08-27 | Debut: complete measured data for four cards across nine game scenes, full-dimension Transformer gen 2 quality/performance/latency comparison |
| 2026-08-28 | Planned additions: ray-tracing tests of Starfield’s Shattered Space DLC new scenes, and launch-day tests of Indiana Jones and the Great Circle |
| 2026-09-XX | Planned additions: RTX 5090 hands-on testing, in-depth 6x MFG real-world latency tests |