ComfyUI master · merged 2026-11-05 · not an upscaler here
RelayVSR in ComfyUI: x2/x4 upscale notes, not an online tool
Pull request 17511, Support RelayVSR, was merged into ComfyUI master on 2026-11-05 at 21:14 UTC, commit 2f8943324e1eae89a5ee6c7db93602299246b5f1. It upscales video x2 or x4 with a one-step Wan 2.2 5B on every 15th frame and a Flash decoder for the other frames. This site does not run that graph.
Nightly notice: this is not an official ComfyUI or RelayVSR build, and it is not an upscaler. The change is on ComfyUI master after 2026-11-05. The v0.44.2 backport pull request 17583 was closed on 2026-11-06 without merging. A desktop app built from an older stable tag will not show the nodes. ComfyUI's own README says commits outside stable tags may be very unstable.
A new still, not an upscale
Nano Banana 2.1 makes a picture. It does not upscale video.
Nano AI does not run RelayVSR, RTX VSR, FlashVSR, or SeedVR2. Nano Banana 2.1 costs 8 credits for one 1K image. Eligible new accounts receive 64 credits. The picture is a new still, not a larger copy of a video you already made.
- Made by
- Comfy-Org
- Released
- November 5, 2026
- Run on Nano AI
- Nano Banana 2.1 · 8 credits per image
- Page updated
- November 7, 2026
Source: github.com
Repository cards, not upscaled frames
Downloaded on 2026-11-07. The GitHub counters are GitHub's at that download. Comparison videos from the two X posts were not saved. These cards are not an upscale and not a Nano AI file.
RelayVSR upscales a video you already have
The point of the search is to enlarge an AI video without generating it again. The pull request describes that job: low-quality frames go in, RelayVSRConditioning builds the positive conditioning and an empty latent without a text encoder, and VSRDecode writes frames, with a detail gain on the decoder residual. The Hugging Face card lists an x2 file whose name includes lightx2v and an x4 file. Wildmind's post 2118493695100289374, on 2026-11-06, says real-time streaming video super-resolution is in ComfyUI and names relayvsr x4 and relayvsr x2 lightx2v. The quoted post on the same account says 30 FPS at 1080p and 14 GB. Those round numbers are not the project page's 29.29 FPS and 13.82 GB. Both are cited. They are not merged into one figure.
Project page, 1080p, one NVIDIA A100 80GB, dual-endpoint, 15-frame interval: RelayVSR 29.29 FPS, 13.82 GB peak, 0.327 s first-frame latency. FlashVSR-Tiny on the same line: 7.80 FPS, 24.447 GB, 2.83 s. BennyDaBall post 2118472181747220682, 2026-11-05 22:35 UTC, one RTX 5090, warm, median of 3, 124 frames, 1344x768 to 2688x1536: RelayVSR 5.8 s and 11.6 GB, RTX VSR 1.3 s and 2.9 GB. Detail over a plain 2x resize: 4.8x vs 2.7x. Flicker vs the draft: 0.98 vs 1.18. The post said the PR was not merged yet. SeedVR2 is named in the ComfyUI README under 3D and vision, and the pull request mentions SeedVR2 only as a neighboring code path. No FPS or memory figure for SeedVR2, and no figure for a FlashVSR larger than FlashVSR-Tiny, was on the pages read. RTX Video Super Resolution in the post is NVIDIA's feed-forward path via the comfyui-rtx-simple node, on that one machine. This page does not rerun any of those tests.
RelayVSR at a glance
- Merge
- 2026-11-05 21:14 UTC. Commit 2f8943324e1eae89a5ee6c7db93602299246b5f1. Pull request 17511.
- Backport
- Pull request 17583 for v0.44.2 closed 2026-11-06 without merging.
- Scales
- x2 and x4. The x2 filename includes lightx2v.
- Authors' timing
- Project page, 1080p, one NVIDIA A100 80GB, dual-endpoint, 15-frame interval: RelayVSR 29.29 FPS, 13.82 GB peak, 0.327 s first-frame latency. FlashVSR-Tiny on the same line: 7.80 FPS, 24.447 GB, 2.83 s.
- One 5090 post
- BennyDaBall post 2118472181747220682, 2026-11-05 22:35 UTC, one RTX 5090, warm, median of 3, 124 frames, 1344x768 to 2688x1536: RelayVSR 5.8 s and 11.6 GB, RTX VSR 1.3 s and 2.9 GB. Detail over a plain 2x resize: 4.8x vs 2.7x. Flicker vs the draft: 0.98 vs 1.18. The post said the PR was not merged yet.
- SeedVR2 and full FlashVSR
- No FPS or memory figure on the pages read 2026-11-07.
- This site
- Nano Banana 2.1, 8 credits, 1K. No x2. No x4. No upload.
- Bytes
- 5163295248, 5164179952, 127759464, and 127759784 bytes on the Hugging Face file listing read 2026-11-07.
Native ComfyUI install, then the four files
The commands are the ones printed in the ComfyUI README on master, read 2026-11-07. The README says regular users should use the desktop app, and that the manual line is the words Git clone this repo, without printing a one-line clone command. File names and folders are from the Comfy-Org/RelayVSR model card.
- 1
Prefer the desktop app unless you want master
The README points regular users to https://www.comfy.org/download for Windows and macOS. Desktop builds use the latest stable core. RelayVSR is on master. The closed backport means a stable desktop may not include it yet.
- 2
Or install with comfy-cli
The README prints these two lines. They install ComfyUI. They do not download the RelayVSR weights by themselves.
pip install comfy-cli comfy install - 3
Manual NVIDIA line, then requirements
Python 3.13 is called very well supported. The stable NVIDIA command is below. Nightly PyTorch is a different index, nightly/cu132, and is optional. Then install requirements inside the ComfyUI folder. The README's manual section says Git clone this repo and does not print the URL on that line.
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130 pip install -r requirements.txt - 4
Put the four files in the folders the card names
Two diffusion files go in models/diffusion_models. Two VAE files go in models/vae. The pull request says the nodes are RelayVSRConditioning and VSRDecode, that no text encoder is needed, and that comfy-kitchen window_attention from pull request 302 is required. A tagged comfy-kitchen version that contains 302 was not confirmed on 2026-11-07.
Files on the Comfy-Org card
5163295248, 5164179952, 127759464, and 127759784 bytes on the Hugging Face file listing read 2026-11-07. Original card text points at huggingface.co/kopper/RelayVSR, which had no model card when opened the same day.
| File | Folder | Bytes | Role |
|---|---|---|---|
| relayvsr_x2_lightx2v_int8_convrot.safetensors | models/diffusion_models | 5163295248 | x2 |
| relayvsr_x4_int8_convrot.safetensors | models/diffusion_models | 5164179952 | x4 |
| relayvsr_flash_vae_x2_fp32.safetensors | models/vae | 127759464 | x2 Flash |
| relayvsr_flash_vae_x4_fp32.safetensors | models/vae | 127759784 | x4 Flash |
What the public tests actually are
The paper page is one machine
Project page, 1080p, one NVIDIA A100 80GB, dual-endpoint, 15-frame interval: RelayVSR 29.29 FPS, 13.82 GB peak, 0.327 s first-frame latency. FlashVSR-Tiny on the same line: 7.80 FPS, 24.447 GB, 2.83 s.
The 5090 post is another test
BennyDaBall post 2118472181747220682, 2026-11-05 22:35 UTC, one RTX 5090, warm, median of 3, 124 frames, 1344x768 to 2688x1536: RelayVSR 5.8 s and 11.6 GB, RTX VSR 1.3 s and 2.9 GB. Detail over a plain 2x resize: 4.8x vs 2.7x. Flicker vs the draft: 0.98 vs 1.18. The post said the PR was not merged yet. Face-match numbers in that post are not repeated here, and this site does not score faces.
Missing cells stay empty
SeedVR2 and a FlashVSR larger than Tiny have no number from these sources. The table says so instead of guessing.
No online upscale
Nano Banana 2.1 is 8 credits at 1K. It does not take a video and return a larger one. /replicate-nightmareai is the still-upscaler note. /viggle-remake is a different local video edit.
Public numbers only
Each cell names its source. A blank source is written as not reported. The A100 row and the 5090 row are different tests.
| RelayVSR | RTX VSR | FlashVSR-Tiny | SeedVR2 | |
|---|---|---|---|---|
| What it is | Video upscale in ComfyUI master after 2026-11-05 | NVIDIA feed-forward path in one post | The small model on the author page | Named in the ComfyUI README. No benchmark here. |
| A100 80GB, 1080p | 29.29 FPS, 13.82 GB, 0.327 s | Not in that comparison | 7.80 FPS, 24.447 GB, 2.83 s | Not reported |
| One 5090, 124 frames | 5.8 s, 11.6 GB | 1.3 s, 2.9 GB | Not in that post | Not reported |
| Scale | x2 and x4 on the model card | x2 in that post's output size | Not stated as x2 or x4 on the project line | Not reported |
| Where it runs | Local ComfyUI. Not this website. | The comfyui-rtx-simple node, in that post | The author comparison only | Listed as a native model. No speed here. |
Pages that were opened
How to leave without an upscale
- 1
Install ComfyUI yourself if you want the nodes
Use the README commands above. A stable desktop from before the merge will not show RelayVSRConditioning or VSRDecode.
- 2
Open Nano Banana 2.1 only for a new still
8 credits at 1K. It does not accept a video and it does not offer x2 or x4.
- 3
Keep the two benchmarks apart
Quote 29.29 FPS with the A100 line. Quote 5.8 seconds with the 5090 line. Do not average them.
Stills, if you use the button
They do not describe an upscale, a face, or a video you already rendered.
Cup
A matte ceramic cup on linen, morning window light. no people, no faces, no logos, no readable text.
Courtyard
An empty stone courtyard after rain, one clay pot by a wall. no people, no faces, no logos, no readable text.
Shelf
A folded grey cloth and a brass key on a dark shelf. no people, no faces, no logos, no readable text.
Window
One unbranded plant on a windowsill, soft morning light, empty room. no people, no faces, no logos, no readable text.
Easy mistakes
- The nodes are missing
- You are probably on a stable tag from before the merge, or the closed v0.44.2 backport. Master after commit 2f89433 is the line that contains the pull request.
- Mixing the two timing numbers
- 29.29 FPS is the authors' A100 figure. 5.8 seconds is one person's 5090 figure for 124 frames. They are not the same test.
- Asking this site to enlarge a video
- There is no upload and no x2 or x4 control. The button makes a new still.
A new still, not an upscale
Nano Banana 2.1 makes a picture. It does not upscale video.
Nano AI does not run RelayVSR, RTX VSR, FlashVSR, or SeedVR2. Nano Banana 2.1 costs 8 credits for one 1K image. Eligible new accounts receive 64 credits. The picture is a new still, not a larger copy of a video you already made.
FAQ
RelayVSR questions
What the pull request, the project page, and the two posts said, and what this site will not run.
Can I upscale a video on this page?
No. There is no upload and no x2 or x4 control. The button opens Nano Banana 2.1 for a new still at 8 credits and 1K.
Is it only a nightly build?
It is on master, merged 2026-11-05. The v0.44.2 backport was closed without merging, so a stable desktop can still be behind that commit. The yellow bar is that warning, not a claim that the merge did not happen.
Which number is the official speed?
The authors print 29.29 FPS, 13.82 GB, and 0.327 s on an A100 80GB at 1080p. The 5090 post prints 5.8 s and 11.6 GB for 124 frames. A quoted post prints 30 FPS and 14 GB. Keep them separate.
Why is the post's merge line different?
GitHub shows the merge at 21:14 UTC. The post at 22:35 UTC still says the pull request was not merged yet. Both times are reported. The merge state used here is GitHub's.
Where are the FlashVSR and SeedVR2 numbers?
FlashVSR-Tiny is on the author page: 7.80 FPS, 24.447 GB, 2.83 s. A larger FlashVSR and SeedVR2 had no FPS or memory figure on the pages read, so those cells say not reported.
Is Nano AI ComfyUI or NVIDIA?
No. Nano AI is not affiliated with Comfy-Org, the RelayVSR authors, or NVIDIA. /replicate-nightmareai and /viggle-remake are linked because they are nearby searches, not because this page runs them.
More from Nano AI
Related pages
Nano AI is not affiliated with Comfy-Org, the RelayVSR authors, or NVIDIA. Figures are from the pull request, the project page, the model card, the README, and two public posts as read on 2026-11-07. This page does not upscale video.