Batch Images (Yogurt Nodes)

ComfyUI gets its speed from batching - one IMAGE tensor carrying N frames through a sampler is dramatically cheaper than running the same sampler N times. But building that batch is usually the awkward part, because the moment your images have different resolutions, the batch is illegal. Batch Images solves the boring half: it takes up to 32 image inputs, makes them uniform, and combines them into a single tensor. Then it hands you the count, width, height, and channels as plain numbers, which is a sneaky-useful bonus.

How it works

Every image input (images1 through images32) gets resized so the whole batch shares one resolution, then stacked into a single (N, C, H, W) tensor. How the resize happens is up to you:

• method - stretch (just distort to the target size), fill / crop (scale to cover, then crop the overflow), or pad (scale to fit, pad the leftover with pad_value, default 1 = white). pad is the one that won't wreck your images, which is why it's the default.
• interpolation - nearest, bilinear, bicubic, area, nearest-exact, lanczos. lanczos is the default and the right choice for most upscaling-flavored batching; nearest only when you want pixel-crisp no-smoothing behavior.

Then the whole batch gets sliced with start_index, end_index, and step - literally Python slice semantics ([start:end:step]). end_index of 0 means "to the end," and negative values work like Python's. So you can not only combine 32 images, you can feed in a big batch and pull out a subset, or every-other-frame it with step 2. That makes the node useful even with a single input: it's a batch slicer that also reports its own dimensions.

Inputs that matter

• images1 … images32 - the images to combine. All optional; the node works with one, and the max is thirty-two.
• method / pad_value - how to reconcile resolution differences. Leave on pad unless you have a reason not to.
• start_index / end_index / step - the slice. Leave at 0 / 0 / 1 for "everything," or slice a subset out of a batch you already have.

Outputs

• images - the combined (or sliced) batch. Wire it into a sampler, a vision node, or a save node.
• count - number of images in the output.
• width, height, channels - the dimensions. These are handy for logic: gate a workflow on image count, or feed the resolution into a resize node. The fact that the node reports them is genuinely useful in automation, not just trivia.

Install

Standard Yogurt Nodes install. ComfyUI Manager → search "ComfyUI-YogurtNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes.git
cd ComfyUI-YogurtNodes && pip install -r requirements.txt

Then restart ComfyUI. Pure tensor work - no models, no downloads.

Where people get burned

Two things. First: pad pads with a solid color, and pad_value 1 is white - if your images have wildly different aspect ratios, padding leaves big white bars that downstream nodes see as content. For photo work you often want fill / crop instead, and for grids of same-size images the whole question disappears. Second, remember that slicing with step reorders nothing and drops nothing silently - it follows Python slicing exactly, so start_index 1 means you lose the first image. If your output count mysteriously looks wrong, walk the slice settings before suspecting a bug. And the one genuinely sharp edge: a batch is only as uniform as you make it, so if you combine images of very different resolutions, the resize step is doing real work - keep an eye on what method you chose or your "no distortion" batch suddenly has squished faces.