ReCaVSR video upscaler: ComfyUI workflows for 8, 12 and 16 GB VRAM

Three versions of one ComfyUI workflow for ReCaVSR, a one-step video super-resolution model (built on Wan 2.2) that upscales any video: AI video (MiniMax H3 and others), anime, 3D or real footage. Load your video, press Run. That's it.

Each workflow reads your video's size and picks the biggest output your card can hold (up to 4x, up to 4K). You never have to set the scale yourself, and it stays inside your card's memory.

File For 720p video becomes
recavsr_upscale_8gb_vram.json 8 GB cards (also 10 and 11 GB) 1920x1080 (Full HD)
recavsr_upscale_12gb_vram.json 12 GB cards 2560x1440 (2K)
recavsr_upscale_16gb_vram.json 16 GB cards and bigger 3840x2160 (4K)

Same model, same quality per pixel in all three. A bigger card only allows a bigger output. Pick the file that matches your card. Don't use a smaller card's file on a bigger card: the size picking is worked out for that card's memory.

Output sizes

Your video 8 GB 12 GB 16 GB
480p 854x480 1922x1080 2562x1440 2776x1560
720p 1280x720 1920x1080 2560x1440 3840x2160
MiniMax H3 1248x704 1872x1056 2496x1408 3744x2112
720p vertical 720x1280 1080x1920 1440x2560 2160x3840
1080p 1920x1080 1920x1080 (cleaned, same size) 3360x1890 3840x2160
1080p vertical 1080x1920 1080x1920 (cleaned, same size) 1890x3360 2160x3840

For other shapes, the two boxes Output width and Output height in the workflow show what you will get.

Setup

1. Install the custom node with ComfyUI Manager (search "ReCaVSR"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/PedroMarinhoDev/ComfyUI-ReCaVSR

2. Download 3 model files into ComfyUI/models/recavsr/ (the same files for all three workflows):

File Size Download
transformer_int8_convrot.safetensors 6.3 GB PedroMarinhoDev/ReCaVSR-ComfyUI
prompt.safetensors 4 MB kopper/ReCaVSR
flashdecoder.safetensors 229 MB kopper/ReCaVSR

3. Drag the workflow for your card into ComfyUI, pick your video in Load Video, press Run. The upscaled video is saved in ComfyUI/output/recavsr/, with the original audio and frame rate.

What you need

  • GPU: NVIDIA, RTX 20 series or newer (RTX 30 series or newer for the fastest kernels).
  • ComfyUI: a recent version (0.44 or newer). The workflow uses only ComfyUI's built-in nodes plus ReCaVSR.
  • System RAM: 16 GB minimum for the 8 GB version, 24 GB+ for 12 and 16 GB (32 GB recommended for 4K). Part of the work is kept in RAM so the output fits the card.
  • Disk: about 7 GB for the models.

How it works

  • Auto size (the grey group): reads the first frames of your video, takes its width and height, and works out the scale. It uses the memory numbers from the ReCaVSR node itself, with a safety margin, and only uses sizes with even width and height so the video encoder accepts them.
  • kv_offload is on in all three: the attention history is kept in RAM instead of VRAM. The output is bit-for-bit the same; it only makes room on the card for a bigger output.
  • Long videos are fine. The video is upscaled 8 frames at a time and written to the file as it goes, so a longer clip costs time and disk space, not VRAM.

Tested

All three tested on an RTX 5070 Ti (16 GB) with 24 GB of system RAM. For 8 GB and 12 GB, a dummy program held the rest of the card's memory and ComfyUI was told the card's total was 8 or 12 GB, so ReCaVSR behaved exactly as on a real 8 or 12 GB card. In both tests 2.4 GB of VRAM was used by the desktop (browser, Discord ...), more than a typical setup, and nothing ran out of memory.

Version Test video Output Time
8 GB 1376x768 2064x1152 5 s clip in 49 s
12 GB 1080x1920 (vertical) 1890x3360 5 s clip in 80 s
16 GB 720p 3840x2160 (4K) 15 s clip in about 3 min

Real 8 GB and 12 GB cards (RTX 4060, 3060, 4070 ...) have slower chips than the test card, so expect longer times. The memory limit itself was tested.

Tips

  • Close other apps that use the GPU before you run (games, OBS preview, video in the browser), especially on 8 GB.
  • Colours drift from the original? Set color_fix to wavelet in the ReCaVSR Upscale node.
  • Scene cuts: ReCaVSR treats a clip as one shot, so you may see a short shimmer right after a hard cut. For the cleanest result, upscale each shot as its own clip.
  • Out of memory anyway? Use the workflow for the next smaller card.

Credits

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support