How DLSS 4.5 Upscales Video
NVIDIA's DLSS 4.5 brings a second-generation transformer model to video upscaling, promising sharper detail and fewer artifacts when stretching low-resolution footage. For users wondering whether this AI can turn 720p clips into clean 4K, the short answer is yes, with caveats. The technology analyzes frames intelligently, reconstructing textures and edges rather than simply interpolating pixels. Early demonstrations show it handling extremely degraded sources, including 240p and 360p content, making them surprisingly watchable at higher resolutions.
Also worth reading: How Does AI Video Upscaling to 4K Work and Which Tools Lead in 2026? · What Are the Best FTZ II 4K Settings for AI Video Upscaling to 4K? · Which AI upscaling presets should you use to upscale video and games to 4K?
In practice, DLSS 4.5 and RTX Video deliver impressive results for AI-generated videos and older clips, though "clean" remains relative. While the upscaler removes compression noise and smooths jagged edges, it may invent details that were never present in the original 720p source. For creators seeking 4K output from modest footage, the results are often striking, but purists may notice the difference between true native 4K and AI-reconstructed frames. Still, for most viewers, the improvement is dramatic and immediate.
From 720p to 4K
DLSS 4.5's second-generation Transformer model has raised the bar for what AI upscaling can do, and the question on many creators' minds is whether it can genuinely turn 720p footage into clean 4K. The short answer is yes, with caveats. NVIDIA's latest Super Resolution technology, originally built for games, now powers RTX Video workflows that can take low-resolution AI-generated clips and push them to 4K with impressive detail reconstruction. PCMag even reported gamers using DLSS 4.5 to make 240p "pixel junk" playable, which suggests 720p source material is comfortably within the model's capabilities. Fine textures, faces, and motion edges benefit most, though heavily compressed or artifact-laden sources can still confuse the model.
For practical results, expect the best outcomes from 720p footage with decent bitrate and moderate motion. The Transformer architecture handles temporal consistency better than earlier CNN-based approaches, reducing flicker and shimmer across frames. That said, upscaling adds plausible detail rather than recovering true information, so extremely soft or noisy sources will look enhanced rather than restored. For most 720p clips, though, the jump to convincing 4K is now genuinely achievable on RTX hardware.
Transformer Model and Frame Generation
NVIDIA's DLSS 4.5 introduces a second-generation transformer model for super resolution alongside six-times dynamic multi-frame generation, and early demonstrations suggest it can perform startling feats of reconstruction. Gamers have already shown the technology taking 240p pixel art and rendering it not merely recognizable but genuinely playable, which naturally raises the question of whether ordinary 720p video clips can be pushed to clean 4K. Because the new neural renderer analyzes motion and depth across multiple frames rather than simply sharpening individual pixels, it can recover detail that older upscalers would have blurred or invented poorly. For real-time gameplay and AI-generated footage processed through RTX Video, the results look increasingly convincing.
However, turning 720p clips into clean 4K for archival or cinematic purposes remains a different challenge. DLSS 4.5 is optimized for interactive rendering, where temporal artifacts are masked by motion and the source is synthetic or game-engine geometry. Compressed 720p video carries block noise, banding, and lost high-frequency texture that no transformer can truly invent. The upscale may look sharper and more stable, but calling it clean 4K stretches the claim.
Quality Tests and Real Gameplay
NVIDIA’s DLSS 4.5 introduces a second-generation transformer model that pushes AI upscaling far beyond simple edge sharpening. While the technology was built for real-time gameplay, early tests show it can transform extremely low-resolution inputs—some gamers have made 240p footage playable—into something resembling high-definition output. For 720p video clips, the leap to 4K seems plausible on paper, especially since NVIDIA’s RTX Video already handles local and cloud-based upscaling for prerecorded content. The key question is whether the AI hallucinates too much detail or preserves the original intent of the footage.
In practice, quality varies by source material. Gameplay rendered at 720p tends to upscale cleanly because DLSS recognizes geometric patterns and motion vectors, producing stable 4K images at high frame rates. Pure video clips, however, lack that structured data, so the model must infer textures and edges from limited information. Critics worry this leads to an overly smooth, artificial look—what some call “slop”—yet the results remain impressive for archival or low-bitrate content. For most users, DLSS 4.5 can deliver watchable 4K from 720p, though purists may still prefer native resolution.
Best Settings for AI Upscaling
DLSS 4.5's second-generation transformer model has raised the bar for what's possible with low-resolution footage, and the short answer is yes, with caveats. NVIDIA's latest Super Resolution model can take 720p clips and produce genuinely clean 4K output, especially when the source material is sharp and well-lit to begin with. The transformer architecture excels at reconstructing fine detail like hair, foliage, and textures that earlier models tended to smear or hallucinate. TweakTown's testing of RTX Video on AI-generated footage showed impressive results moving from 720p to 4K, and PCMag reported gamers successfully upscaling even 240p pixel junk into playable content, which suggests 720p sources have far more headroom.
That said, results depend heavily on your settings and source quality. Start with the highest quality preset your GPU can handle, keep temporal stability enabled to reduce flickering between frames, and avoid pushing multi-frame generation too aggressively on video content, since it's designed primarily for gaming frame rates. Compression artifacts, heavy motion blur, and noisy footage will limit how clean the final 4K looks, so feed the model the best possible input.
DLSS 4.5 vs Traditional Upscaling
| Capability | DLSS 4.5 | Traditional Upscaling |
|---|---|---|
| 720p to 4K Quality | Clean, sharp detail via transformer AI | Soft, artifact-prone output |
| Temporal Coherence | 6X Multi Frame Generation smooths motion | Independent frames cause flicker |
| Low-Rescue Ability | Makes 240p sources playable | Struggles below 480p |
| Real-Time Performance | Optimized for RTX GPUs | CPU-heavy, slower results |