news

FlashVSR+ claims untiled 4K video upscaling on one 24GB RTX 4090, trained against a black-box reward model

The paper reports up to 2x DiT speedup and 45% lower peak memory, with preference alignment done through reward-weighted teacher rollouts instead of backpropagating through the reward model.

By Priya Shenoy · 16 October 2026

A research team has published a video upscaler that claims 4K output on a single consumer graphics card without splitting the frame into tiles, and the training method behind it is as relevant to this site's beat as the speed numbers. FlashVSR+, from Shanghai AI Lab, MMLab CUHK, Tsinghua University and CPII under InnoHK, went up on arXiv as paper 2610.17552 with code on GitHub, and it aligns a distilled one-step model using a reward model it never backpropagates through.

The reported numbers: 2x DiT speedup, 45% less peak memory

The architecture replaces most dense attention blocks with local spatiotemporal attention while keeping a small number of sparse global blocks for long-range consistency, a design the authors justify by observing that attention in pretrained video super-resolution backbones is concentrated in compact neighborhoods. The paper reports up to 2x DiT speedup and 45% lower peak memory against dense attention, which is what makes spatially untiled 4K inference fit on a single 24GB RTX 4090. Tiling is the usual way to get 4K out of a consumer card, and it is also the usual source of boundary artifacts, so removing it is the substantive claim here.

Ry-DMD puts a non-differentiable judge inside the training loop

The second half of the paper is Reward-Yield Distribution Matching Distillation, which the authors describe as using reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model. In plain terms, the scoring function does not have to be differentiable or even inspectable, it just has to rank samples. The paper states the method generalizes beyond video super-resolution to text-to-image generation, which is where this stops being a niche restoration result and starts being a general recipe for training generators against whatever evaluator you happen to trust.

That matters because the evaluator becomes the specification. If the reward model is an opaque preference scorer, the output distribution inherits its blind spots, and nothing in the training signal flags where those are. Legibility of rendered text is one of the places that shows up fastest in practice, which is why we score it as its own category across models including Seedance 2.5.

What the release does not include

Every figure above is self-reported by the authors, with no third-party evaluation attached at publication. The GitHub repository also narrows the hardware claim: the released inference pipeline is Linux and CUDA only, does not support CPU inference, and requires input height and width both divisible by 128. The project ships a custom Triton attention backend that the authors recommend specifically for the RTX 4090 and other consumer cards where their alternative backend may be unavailable, so the headline single-GPU result depends on using their kernel rather than a stock setup.

Upscaling is where a lot of the 4K in this market actually comes from

Most consumer AI video pipelines generate below 4K and reach it through a separate enhancement pass, which means the cost and the artifacts of that pass are a real part of what a 4K claim is worth. An upscaler that runs untiled on a desktop card moves that step off a rented GPU, and it moves the quality decision onto whatever reward model was used to tune it.

Sources: arXiv, OpenImagingLab on GitHub.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.