Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)

RelayVSR upscales video by 4× and delivers 29 fps on an A100

Making old video sharper is appealing, but running every frame through a heavy neural network is expensive. RelayVSR splits the workload: a large model prepares reference frames, and a small model uses them to process the entire clip at 4× resolution. The authors achieved 29.29 fps at 1080p output on an NVIDIA A100; developers should first test speed and quality on their own footage.

AI-processed from TG: @neuraldvig (Нейродвиж); edited by Hamidun News
RelayVSR upscales video by 4× and delivers 29 fps on an A100
Source: TG: @neuraldvig (Нейродвиж). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @neuraldvig (Нейродвиж) ↗
Editor-in-Chief
Key metric #1Peak memory usage in the test was 13.82 GB, but speed was measured on an A100.
Key metric #2Code, weights and instructions for running on Linux with NVIDIA hardware are available.
Audio must be restored separatelythe current script does not carry it over.

An old clip needs to go on a big screen, but conventional upscaling leaves faces blurry and turns detail into mush. Generative upscaling can produce a more convincing image, but processing every frame with a heavy model is expensive. RelayVSR proposes splitting the workload: a large network prepares occasional reference frames, while a small one processes the rest of the sequence. For a video product team, this is a reason to test local processing on its own clips and calculate the cost per minute.

Why use two models to process one clip

Imagine footage of someone wearing a textured jacket. The large model creates a detailed reference for selected moments, and the lightweight network uses it alongside the source video to process the intermediate frames. It does not have to invent the entire fabric texture from scratch each time. It also takes recent frames into account to keep details consistent with the motion.

This division of labor has an unwelcome property: an error in a reference frame can propagate to subsequent frames. That is why the authors train the large model with the quality of the final video in mind. Judging the visual appeal of an individual frame is not enough: a good image also needs to help the second network assemble a coherent video. The method is described in the RelayVSR paper.

The practical principle extends beyond upscaling. If an expensive operation can be performed occasionally and its output reused, there is an opportunity to reduce computation. But savings must be assessed alongside errors: one poor reference can spoil several subsequent outputs. For a video service, this affects both costs and the share of footage that needs to be processed again.

What the 29 fps result means

According to the authors, RelayVSR reaches 29.29 fps at 1080p output resolution, with peak GPU memory usage of 13.82 GB. In the same comparison, FlashVSR-Tiny delivers 7.80 fps and uses 24.447 GB. The measurement was performed on a single NVIDIA A100 with 80 GB of memory, with reference frames spaced 15 frames apart. These are the project team's results, published on the demo page.

These results do not support the claim in a Telegram post that any graphics card with more than 14 GB of memory will work. Memory capacity describes only part of the requirements for running the tool. It guarantees neither hardware compatibility nor the same speed on a home computer. Buying a card for this tool based on a single number in a post would be premature.

There is also the question of how much to trust the image. A more convincing texture does not, by itself, prove that this was the texture in front of the camera. For a family archive, that may be an acceptable compromise. For footage where the accuracy of fine details matters, the source should be preserved and compared with the output. The promise to “rescue any video” is too broad: the quality of each recording still needs to be assessed visually.

How to assess its value before integration

The official repository provides instructions and links to two weight files. The documented runtime environment is Linux, Python 3.11–3.13 and an NVIDIA GPU. The script increases both width and height by exactly four times and preserves the frame rate, but does not carry over audio. A production workflow will therefore need a separate step to restore the audio track.

A sensible starting point is short excerpts of your own footage: a face in motion, fine textures and an abrupt scene change. Check whether facial features are preserved, whether details start to flicker and how long processing takes. These checks will do more to inform an integration decision than an impressive still frame from a demo.

For a product manager, the next step is to measure the full processing cost, including file preparation, audio restoration and repeat runs for scenes that fail. For a developer, it is to check reproducibility on the available hardware. RelayVSR already provides enough for that experiment: a published implementation, weights and a specific set of instructions for running it. The decision to adopt it should follow testing on the video the product will actually handle.

Story

Key Questions Answered

Why use two models to process one clip?

Imagine footage of someone wearing a textured jacket. The large model creates a detailed reference for selected moments, and the lightweight network uses it alongside the source video to process the intermediate frames. It does not have to invent the entire fabric texture from scratch each time. It also takes recent frames into account to keep details consistent with the motion. This division of labor has an unwelcome property: an error in a reference frame can propagate to subsequent frames.

What the 29 fps result means?

According to the authors, RelayVSR reaches 29.29 fps at 1080p output resolution, with peak GPU memory usage of 13.82 GB. In the same comparison, FlashVSR-Tiny delivers 7.80 fps and uses 24.447 GB. The measurement was performed on a single NVIDIA A100 with 80 GB of memory, with reference frames spaced 15 frames apart. These are the project team's results, published on the demo page. These results do not support the claim in a Telegram post that any graphics card with more than 14 GB of memory will work.

How to assess its value before integration?

The official repository provides instructions and links to two weight files. The documented runtime environment is Linux, Python 3.11–3.13 and an NVIDIA GPU. The script increases both width and height by exactly four times and preserves the frame rate, but does not carry over audio. A production workflow will therefore need a separate step to restore the audio track. A sensible starting point is short excerpts of your own footage: a face in motion, fine textures and an abrupt scene change.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…