The Pulse
FlashVSR+ Targets Untiled 4K Video on an RTX 4090
FlashVSR+’s authors report spatially untiled 4K video super-resolution on a 24GB RTX 4090, using hybrid attention and CPU memory for part of the cache. Their paper also describes reward-guided training and reports results from a user study
AI.info Team · Written with AI ·
FlashVSR+ makes a demanding claim about 4K video restoration: it can process each frame at full spatial resolution on a 24GB RTX 4090, without dividing the image into tiles. But fitting that workload on a consumer GPU requires more than an attention redesign—the system also moves parts of its memory workload between the GPU and host RAM. The researchers’ October 13 paper pairs that efficiency work with a second aim: steering a fast, one-step diffusion model toward outputs people prefer.
“It achieves up to 2× DiT speedup and 45% lower peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090.”
Ruchang Yao and co-authors, the paper’s authors, write in FlashVSR+.
Full-frame 4K, with a memory trade-off
Video super-resolution (VSR) reconstructs higher-resolution video from lower-resolution input. FlashVSR+ uses a one-step diffusion model, aiming to deliver restoration without the repeated inference steps that make many diffusion systems slow. Its hybrid attention design concentrates computation on nearby space-time blocks, where the authors say pretrained VSR models place much of their attention, while preserving sparse global attention for consistency across more distant parts of a clip.
The paper’s RTX 4090 comparison uses 89-frame videos and includes CPU key-value-cache offloading and asynchronous prefetching overhead. That distinction matters: “untiled” means the model processes each frame at full spatial resolution rather than chopping it into image patches; it does not mean every piece of the workload stays in GPU memory. For comparison, the authors say SeedVR2 runs out of memory with full-frame inference at 2K and 4K, so their benchmark uses tiled inference for that model.
Turbo runs faster; Base leads the preference test
The researchers offer Base and Turbo versions. The project’s 4K RTX 4090 benchmark lists throughput of 0.50 frames per second for Base and 0.98 for Turbo, against 0.18 for SeedVR2. The project describes Turbo as the faster option and Base as the higher-quality choice; those results are the team’s own benchmark figures, not an independent evaluation.
The paper also reports a blinded pairwise user study on 20 videos sampled from three benchmarks: RealisVideo-4K, VideoLQ and AIGC35. Base serves as the reference point and scores better than the compared methods on the study’s fidelity, quality and overall-preference measures. Turbo remains close, with an overall preference margin of -6% against Base, illustrating the trade-off between its higher inference efficiency and the team’s preferred restoration output.
Reward-guided training targets legibility
FlashVSR+’s second contribution, Reward-Yield Distribution Matching Distillation, or Ry-DMD, uses reward scores to tilt a teacher model’s training targets toward outputs that score better. The method does not backpropagate through the reward model, which can be a black box. The authors test the approach on video restoration and text-to-image generation.
In an OCR-focused VSR experiment, the paper reports that Ry-DMD raises the OCR score from 0.624 for the baseline to 0.776, while its reported PSNR rises from 23.94 to 24.86. Those results come from a synthetic text-focused dataset, so they show performance on that specific task rather than a general guarantee that restored text will be accurate in everyday footage.
Code and checkpoints are available
The team has published inference code and Base and Turbo checkpoints, alongside a dataset release. The repository specifies Linux, an NVIDIA GPU with CUDA, and Python 3.10; input height and width must both be divisible by 128. Its README recommends a Triton attention backend for the RTX 4090, while listing other backend options for supported hardware.
The paper was submitted to arXiv on October 13, 2026. Its central result is a spatially untiled 4K run on one 24GB RTX 4090, paired with memory management that shifts cache work to the CPU. That combination—not a claim that the entire computation remains on the graphics card—defines what the release demonstrates.