Pt Pad

PtPad pads an image tensor up to a target height and width, centered, with black (zero) borders. It's the pack's answer to "I have a bunch of images of different sizes and I need them all to be the same shape before they go into a model" - the letterboxing step you'd normally do with PIL, done on raw tensors in the graph. This pack's image workflows are built on tensors, and PtPad is the shape-normalization tool that keeps them feeding into fixed-size models.

How it works

The node takes a tensor of shape (c, h, w) or (b, c, h, w), creates a new zero tensor of your target height × width, and centers the original inside it. Three inputs:

• tens - the input tensor.
• height / width - the target dimensions. The padding is centered: extra space is split between top/bottom and left/right.

Output is a single TENSOR of exactly (c, target_h, target_w) or (b, c, target_h, target_w).

The crucial gotcha: this is not a ComfyUI IMAGE

Here's where people lose an afternoon. ComfyUI's IMAGE type is (b, h, w, c) - channels last - but this node expects channels-first tensors. The pack's own docstring spells out the dance: to pad a ComfyUI image, first convert with Pt From Image, then use Pt Permute with (0, 2, 3, 1) to transpose (b, h, w, c) → (b, c, h, w), then pad, then permute back with (0, 3, 1, 2) and convert with Pt To Image. Skip a step and you'll either get a shape error or, worse, a silently transposed image. There's also an explicit guard: if your target height/width is smaller than the input, the node raises a clear error rather than silently cropping - padding can only grow, which is the right behavior, just don't expect it to shrink.

When you'd use it

Batching images of different sizes for a training or inference run, or normalizing a single image to the exact resolution your model expects before prediction. Combined with Pt Permute, it's the standard prep pipeline between "image loaded in ComfyUI" and "tensor fed to a model node."

Installing it

Part of ComfyUI-Pt-Wrapper (HowToSD's no-code PyTorch pack, a spin-off of ComfyUI-Data-Analysis). ComfyUI Manager → search "ComfyUI-Pt-Wrapper", or:

cd ComfyUI/custom_nodes
git clone https://github.com/HowToSD/ComfyUI-Pt-Wrapper

Restart after; first boot is slow while pandas, scikit-learn, transformers, sentencepiece, peft and friends install. No model downloads needed.