CV Stitch (Advanced List)

This is CV Stitch (Advanced) with one substitution: images takes a ComfyUI list instead of a batched NPARRAY. Each item can be an IMAGE, a LATENT or an NPARRAY, and only the first frame of each item is used.

If that sounds like a footnote, it isn't. Lists and batches are different shapes in ComfyUI, and which one you have depends on where your images came from - a folder loader handing you a list of images wants this node, while a batched IMAGE clip wants the batch version (or a trip through Image Batch → CV Batch). Getting it wrong isn't an error, it's a quiet one: feed a batch into a list socket and you stitch a single frame. The example graph builds its list with core's Batch Images node, which is the usual assembly point.

Everything else is the full cv2.detail pipeline rather than the one-call cv2.Stitcher: feature detection, matching, homography or affine estimation, bundle adjustment, warp projection, exposure compensation, seam finding, blending. Stage order and parameter set follow OpenCV's own stitching_detailed.py, so the knobs mean what they mean in the reference implementation.

The knobs that earn their place

warp_type is the big one - spherical by default and right for a wide sweep, but plane is what you want for flat, coplanar subjects (documents, posters, a wall). Spherical on a flat subject is the classic bow-tie distortion.

seam_finder - gc_colorgrad (GraphCut on colour plus gradient) is the best all-rounder; dp_color / dp_colorgrad are cheaper dynamic-programming cuts; voronoi is crude and fast; no disables it. blend_type picks multiband (Laplacian pyramid, best), feather (linear) or no (hard cut), with blend_strength setting how soft the merge is.

exposure_compensation deals with the brightness jumps your camera's metering created between frames: gain_blocks default, gain for one global gain, channel/channel_blocks for colour, no to leave it alone.

Then features (orb / sift), matcher, match_conf (-1 auto), estimator (homography for perspective, affine for planar SCANS-style captures), bundle_adjuster, pano_confidence_thresh, and the three resolution stages - registration_resol, seam_estimation_resol, compositing_resol - all at -1 for full resolution, and the first place to look when memory runs out. wave_correction (auto by default) handles hand-held drift. interpolation_flags sets the warp interpolator.

There's an optional masks list too, one coverage mask per image.

Outputs: panorama (uint8 BGR NPARRAY), status (OK or an ERR_* label) and success. Failures return a copy of the first input frame instead of raising, so branch on success. Stitching happens in a separate process and stays interruptible.

Install

ComfyUI Manager → search ComfyUI CV (publisher bmad4ever), or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv
pip install "opencv-contrib-python-headless~=5.0.0.93"

Restart after. Python ≥ 3.12 and a recent ComfyUI on the V3 node API. Keep the contrib wheel: a plain opencv-python install shares one site-packages/cv2 and silently empties the contrib submodules other nodes here rely on. tools/repair_opencv_contrib.py --check and --apply are in the repo.

48_panorama_playground.json uses this node alongside the batch variants - a genuinely good way to see what each parameter does, since all three run on the same two images. Get those files in place with workflows/01_install_example_inputs.json, then reload the page.

Where it bites

Change one parameter per run. Twenty knobs with visual failure modes is not a place to be brave. And remember the item ordering is the stitch order - a list assembled out of sequence produces a panorama that's technically successful and completely wrong.

The pack caveat applies as always, and it's worth repeating because the author says it himself: this is a personal, heavily LLM-assisted project, updates aren't planned, and the workflows demonstrate functionality rather than ship it. The parameter set here is lifted from OpenCV's official sample, which is the reassuring part; the preset values are one person's taste.