CV QR Encode

Text in, QR bitmap out, all inside the graph. If your workflow needs to stamp a code onto an image - a link, a caption, a serial number, a per-generation seed encoded for later - this is the node that makes it instead of you leaving ComfyUI for a website.

How it works

cv2.QRCodeEncoder writes the symbol, and the node upscales it with nearest-neighbour so the modules stay crisp - a bilinear-upscaled QR is a QR that scans badly, because the finder pattern's hard edges get smeared. module_size is pixels per module: 1 is the raw bitmap (too small for most scanners), 6–12 is the sweet spot for print or screen.

Two knobs deserve actual thought:

error_correction is the Reed-Solomon level - how much of the code can be damaged, occluded, or obscured and still decode. L is about 7% recovery, M 15%, Q 25%, H 30%. Higher levels leave less room for data, so the symbol grows for the same payload. If you're pasting a logo over the centre or printing onto something that gets handled, go up; if you're engraving pixels, M is a sane default.

quiet_zone is the one that catches people out. OpenCV frames the symbol with only two quiet-zone modules per side where the QR spec asks for four. The default of 2 here adds the missing margin, making the output spec-compliant. If you're compositing a code onto a busy background, raise it - scanners need that white ring, and a code with no breathing room fails for reasons that look like nothing is wrong.

version fixes the symbol size (v = 21 + 4·(v−1) modules square, so v1 is 21x21 and v40 is 177x177); leave it at 0 and OpenCV picks the smallest version that fits your payload at the chosen error-correction level. The mode dropdown controls the payload encoding - auto picks the densest mode that fits; numeric and alphanumeric pack more characters per module at the cost of a restricted alphabet; byte carries arbitrary bytes; ECI tags the payload as UTF-8, which is what you want for non-Latin text; kanji packs Shift-JIS.

Structured Append

Set structure_number above 1 and the message is split across that many codes - the output becomes an IMAGE batch of the parts, white-padded to a common size. This is the trick for payloads too long for one symbol, and the parts decode as a single joined message only when a decoder sees all of them; CV QR Detect writes the joined text on the first part.

Inputs and outputs

text is the payload, and it has to be representable in the chosen mode and fit the chosen version - otherwise the node reports OpenCV's error rather than writing a corrupt code, which is the behaviour you want.

Alongside qr_image you get modules, the encoder bitmap's side length in modules including OpenCV's two-module quiet zone but excluding your extra one (so version 1 reads 25), and version, the version actually used - which is how you find out what auto resolved to. Both are useful sanity checks when a long payload quietly bumps you up several versions and your code comes out much denser than planned.

The output is an ordinary IMAGE with three identical channels, so it composites, scales, rotates and saves like any other image. A common pattern: encode, then warp the code into place with CV Quad Warp in fill mode so it sits on a photographed surface.

Installing

ComfyUI Manager → search "ComfyUI CV", or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv
# restart ComfyUI

Python ≥ 3.12, a recent ComfyUI on the V3 node API, and opencv-contrib-python-headless~=5.0.0.93 (plus numpy, torch). No model files - OpenCV's encoder is self-contained, unlike the ONNX-based WeChat detector in the same pack.

Gotchas

Don't upscale the output with a smooth filter. A Lanczos or bicubic resize of a QR adds ringing around module edges. If you must resize, keep it integer-scaled and nearest-neighbour, or generate at the size you need via module_size.

A code that won't scan is usually a margin or contrast problem, not an encoder problem. Check the quiet zone, and make sure the code isn't dark-on-dark or light-on-light after compositing.

Non-Latin text needs the ECI mode. Encode Japanese or Cyrillic with auto and you may get a symbol that decodes to mojibake on a phone.