Skip to content
Bud's Wiki

RTX Video SDK — RTX Video Super Resolution and RTX Video HDR

TL;DR

Section titled “TL;DR”

NVIDIA’s RTX Video SDK bundles two Tensor-Core-accelerated AI video pipelines: RTX Video Super Resolution (upscale + sharpen + compression-artifact removal) and RTX Video HDR (SDR→HDR10 tone mapping). Both run as feed-forward CNN-style models optimized for real-time playback or fast offline rendering on RTX GPUs; alex’s internal test upscaled a 720p Ray3 SDR clip to 4K in ~5 seconds. Not a research artifact — a shipping productized inference SDK with no published architecture details, training data, or benchmark numbers. Notable for Luma because it’s the consumer-grade baseline any internal upscaling/HDR work has to clear on quality and on latency.

Key claims

Section titled “Key claims”
  • RTX Video Super Resolution is an AI model trained to “sharpen edges, restore features and patterns” and remove blocky compression artifacts, targeting 1080p low-bitrate input → 4K output [§RTX Video Super Resolution].
  • Artifact reduction is bundled with upscaling because compression artifacts become more visible after sharpening, so the model has to denoise and upsample jointly [§RTX Video Super Resolution].
  • The model runs on Tensor Cores on NVIDIA RTX GPUs, with two deployment modes: offline processing in creative apps, and real-time playback in media playback apps [§RTX Video Super Resolution].
  • RTX Video HDR expands SDR colorspace content to HDR10 via an AI model trained on paired SDR/HDR frames, with additional tuning parameters available to creators for stylization [§RTX Video HDR].
  • HDR is positioned for creators mixing HDR and SDR source footage in HDR10 projects, not as a general-purpose tone mapper [§RTX Video HDR].
  • The two effects compose: running super resolution + HDR on the same frame produces 4K HDR output from 1080p SDR input in a single pass [§Sharpness and Clarity Improvement in Two Dimensions].

Method

Section titled “Method”

The product page describes neither architecture nor training procedure. From the framing (“AI model … optimized to run using the powerful Tensor Cores”) and the real-time-playback claim, both effects are almost certainly compact convolutional models — not diffusion — quantized/tuned for NVIDIA’s TensorRT-style Tensor Core path. The SDK exposes them as effects that can be inserted into creative or media-playback pipelines; there is no mention of temporal modeling, optical flow, or cross-frame propagation, so the per-frame interpretation is the safe default until NVIDIA documents otherwise.

The internal evidence point in the Slack thread (5s to upscale a 720p→4K Ray3 clip) is consistent with single-pass per-frame inference at high throughput on a recent RTX GPU, not with multi-step diffusion-based VSR.

Results

Section titled “Results”

No quantitative benchmarks are published. The product page advertises three demonstration frame comparisons (SDR vs HDR-enhanced; 1080p SDR vs 4K SR+HDR-enhanced) and the qualitative claim of “sharper, cleaner” output. Internal team result from alex: ~5 seconds for a 720p→4K upscale of a 5-second Ray3 generation. No FPS or quality metric (PSNR/SSIM/LPIPS/MUSIQ/JOD) reported anywhere.

Why it’s interesting

Section titled “Why it’s interesting”

This is the productized counterpoint to the wiki’s Video Super-Resolution / Restoration via Diffusion Priors cluster, which is dominated by 2025-26 diffusion-based VSR (SparkVSR, FlashVSR, Vivid-VR, DUO-VSR). Those papers report perceptual gains via large pretrained T2V backbones (CogVideoX1.5-5B, Wan2.1) at the cost of GPU memory and latency that put real-time playback out of reach without aggressive distillation. RTX Video SDK is the opposite end of the operating curve: a non-diffusion CNN-style model shipped on consumer GPUs, evaluated only by eye but already in the hands of users — exactly the latency × quality envelope FlashVSR v1.1 — Real-Time Diffusion-Based Streaming Video Super-Resolution (HF model release) (~17 FPS at 768×1408 on a single A100 via LCSA + Tiny Decoder) is trying to reach from the diffusion side. The SDR→HDR pairing is also a useful contrast to HDR Video Generation via Latent Alignment with Logarithmic Encoding (LumiVid) (LumiVid) and DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models (DiffHDR), which both adapt a frozen video DiT via LoRA + a logarithmic input transform for HDR reconstruction — a heavyweight approach against RTX’s feed-forward shipped product. Worth tracking what gap remains once a researcher actually compares them head-to-head with calibrated metrics.

See also

Section titled “See also”