Remember when upscaling a short AI video meant waiting several minutes, or sometimes even longer?
Well, things are getting considerably faster.
ComfyUI just added support for RelayVSR, an open-source AI video upscaler capable of processing 1080p footage at nearly 30 frames per second.
And the technology behind it is genuinely interesting.
Meet RelayVSR, a streaming video super-resolution system developed by researchers at the University of Science and Technology of China.
Unlike traditional generative upscalers, it doesn't run a massive diffusion model on every single video frame.
Instead, it makes the expensive AI model work only when necessary.
Here's how it works.
Step 1: Generate detailed reference frames
A large generative model creates high-quality references at selected keyframes.
Step 2: Reconstruct everything between them
A much smaller network called the Dual-Memory Video Transformer transfers visual details from those reference frames to the rest of the video.
It remembers both the generated references and the recent motion context.
Step 3: Keep the video consistent
A reinforcement learning technique called Video-Aware Reference Optimization (VARO) helps prevent visual errors from accumulating across frames.
The result is a system designed to upscale long videos without repeatedly running expensive generative inference on every frame.
And the performance numbers are impressive.
According to the researchers' benchmarks, on a single NVIDIA A100 80GB:
RelayVSR: 29.29 FPS at 1080p
FlashVSR-Tiny: 7.80 FPS
RelayVSR GPU memory: 13.82 GB
FlashVSR-Tiny GPU memory: 24.45 GB
First-frame latency: Just 0.327 seconds
That's approximately 3.75× faster, while using considerably less GPU memory.
But here's an important distinction.
The nearly 30 FPS benchmark was measured on an NVIDIA A100, not a typical consumer graphics card.
And 13.82 GB of VRAM usage doesn't mean every 16 GB GPU can achieve the same speed.
The architecture supports efficient streaming, but performance still depends on hardware, resolution, configuration, and video content.
Now for the interesting part for ComfyUI users.
RelayVSR support has been merged into ComfyUI, and the team has published optimized model weights.
Available configurations include:
4× video upscaling: Higher-resolution reconstruction using the x4 model.
2× video upscaling: A lighter alternative based on LightX2V.
INT8 quantization: Reduced model memory requirements.
Streaming processing: Designed for longer video sequences.
Open-source inference: Run the models locally instead of relying on a commercial upscaling API.
The approach can be used for enhancing low-resolution footage, AI-generated videos, animations, and other visual material.
And yes, 4K output is possible with suitable input resolutions and hardware, although the published near-real-time benchmark is for 1080p.
Want to try it?
Download the ComfyUI-ready weights:
https://huggingface.co/Comfy-Org/RelayVSR
Place the diffusion models and VAE files in their corresponding ComfyUI model directories, update to a build that includes RelayVSR support, and configure your workflow.
But here's what I find most interesting.
We're watching AI video production move toward increasingly practical workflows.
Generating impressive footage is one thing.
But actually using that footage in a production pipeline requires resolution, temporal consistency, predictable performance, and reasonable hardware requirements.
And RelayVSR addresses several of those problems at once.
Instead of making one giant AI model process everything, the researchers combined a powerful model with a much smaller, specialized one.
The big model creates the important details.
The small model handles the repetitive work.
And the entire system becomes dramatically more efficient.
It's a familiar lesson from agentic AI, now applied to video processing.
Maybe the future of generative video isn't just bigger models.
Maybe it's smarter ways to make different models work together.
And getting close to real-time generative upscaling is a pretty significant step in that direction.
RelayVSR models for ComfyUI:
https://huggingface.co/Comfy-Org/RelayVSR
Official research and interactive demonstrations:
https://kopperx.github.io/RelayVSR/
Source code:
https://github.com/kopperx/RelayVSR
Research paper:
https://arxiv.org/abs/2609.37850
Free and open source under Apache 2.0.
#RelayVSR #ComfyUI #AIVideo #VideoUpscaling #SuperResolution #LocalAI #GenerativeAI
YouTube / ABVX