Ptv Transforms Resize

Ptv Transforms Resize builds the image preprocessing step for your training pipeline: it produces a transform that resizes every image to a target height and width and converts it to a PyTorch tensor. In the pack's image-classification flows - like the dog-vs-cat tutorial - this is the node that makes your raw images the right size for the model before they ever reach the dataloader.

It's part of ComfyUI-Pt-Wrapper, the ~200-node pack that brings PyTorch data loading and model training into ComfyUI's graph. The Ptv prefix is the dataset side of the family: these nodes produce PTVTRANSFORM objects, the transform recipes that dataset and dataloader nodes consume. PtvTransformsResize is one of the few transform-building nodes, sitting alongside PtvTransformsToTensor.

How it works

Under the hood it wraps torchvision, the standard PyTorch image tooling:

transforms.Compose([
    transforms.Resize((height, width)),
    transforms.ToTensor()
])

That's the whole mechanism. The output is a single PTVTRANSFORM - a callable recipe, not a transformed image - and you hand it to a dataset node (like PtvImageFolderDataset or the other Ptv dataset nodes). When the dataset loads each image, the transform runs: the image gets resized to exactly height × width, then ToTensor converts it to a C,H,W tensor with values scaled into [0, 1]. One transform node, applied consistently to every image in your training and validation sets - which is exactly what you want, because inconsistent sizes are a classic training error.

Inputs and outputs

Only two knobs:

• height (INT, default 256) - target height in pixels.
• width (INT, default 256) - target width in pixels.
• PTVTRANSFORM output - the resize+tensor transform, ready to wire into a dataset node.

Pick dimensions that match your model's input. The pack's example models (ConvNet, ResNet) have fixed input sizes, so match the height/width here to whatever the model expects - mismatch and you'll get shape errors at the model call.

Installing the pack

Install the pack once, get all ~200 nodes:

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

Restart ComfyUI. Or ComfyUI Manager → search "ComfyUI-Pt-Wrapper" → install. It's on the Comfy Registry, so Manager handles the dependencies (pandas, scikit-learn, transformers, sentencepiece, and the rest). Because this is a training node, you'll also want the pack's example workflows from examples/workflows/ - dragging one in gets you the whole dataset→transform→dataloader→train chain pre-wired.

Common issues

• Wrong input size. The most common failure, and it's a shape mismatch, not a crash you'll clearly understand. The model expects, say, 96×96 but your transform outputs 256×256 - check that height/width match the model node's config.
• Resizing distorts aspect ratios. Resize stretches to the exact box. Fine for training, but if your evaluation metrics drop, a square-crop or pad strategy may serve you better than squashing.
• Transform not applying. The transform only runs through a dataset node that consumes PTVTRANSFORM. If you've wired the output to nothing, your images stay raw.

It's the unglamorous but essential step: every training run that works is one where images showed up at the model in the right size. This node is how you make that true in the graph.