Image grid

If you've ever built a "which checkpoint wins?" workflow, you know the pain: eight images on eight separate preview nodes, then you spend the next ten minutes alt-tabbing between tabs trying to remember which one had the better hands. LF_ImageGrid is the fix - it renders a whole set of images into one labeled comparison matrix so the differences are visible in a single glance.

The layout is where it gets interesting. The grid isn't just "images in a row." The image input takes images in row-major order, and if you feed it a dataset, that dataset's columns become the matrix's columns and its nodes become the rows - with column titles rendered above and node values beside each row. No dataset? You get one automatic row and the images are just numbered. That "dataset" comes from LF_WriteJSON (another LF Nodes node), so you can programmatically drive row/column labels from whatever produced the images.

What it does with your pixels

Every cell is rendered to cell_width × cell_height (both default 512) by fitting - images are scaled and letterboxed, never cropped or stretched. gap_px (default 8) spaces cells and headers, background picks black, white, or transparent (transparent gives you an RGBA result), and title puts an optional heading above the whole thing. show_headers turns the column titles and row value labels on or off.

Because it preserves the original aspect ratio inside each cell, the tool is honest about comparisons: mismatched resolutions show up as letterboxing, not as silently deformed previews. That's a feature, though some people find it odd the first time they see uneven boxes.

Inputs and outputs worth knowing

The inputs you'll actually touch: image, cell_width/cell_height (they also let you decide how big each comparison is - drop to 256 if you're comparing 20 at once), background, and show_headers. dataset only matters when you want named rows/columns, and ui_widget is an internal LF Masonry preview you can ignore.

Outputs are image (the rendered matrix - wire it into a SaveImage or preview), dataset (the same LfDataDataset, now with the composed cells), image_list (the matrix as a one-item list, handy for nodes that only eat lists), and receipt, a deterministic lf.image_grid.receipt.v1 layout record.

Installing and gotchas

It ships in the LF Nodes pack - Manager, search "LF Nodes," or:

cd ComfyUI/custom_nodes
git clone https://github.com/lucafoscili/lf-nodes.git

then restart. Manager auto-installs the pack's requirements.txt (heavier stuff like rembg and onnxruntime is for other nodes in the suite; this one just needs torch and Pillow). Note the pack was rewritten and the old repo archived in 2025 - make sure you're on lucafoscili/lf-nodes.

Two things trip people up. First, the grid is bounded by the image input's ordering - if you don't keep your sources in the exact order you want displayed, the labels won't match the pictures, and there's no drag-to-reorder. Second, cell sizes cap at 4096 and there's a total-pixel safety limit, so a 10×10 grid at 1024 cells each isn't happening - drop the cell size. Otherwise this is one of the least finicky image nodes in the pack: feed it a batch, read the result.