Introduction

In Keras, padding for Conv2D controls what happens at the image borders before convolution is applied. The built-in padding argument is simple by design: you usually choose either "valid" or "same". If you need an exact custom border width, you usually add a separate padding layer such as ZeroPadding2D before the convolution.

Use padding="valid" for No Padding

"valid" means no extra pixels are added.

1from tensorflow import keras
2
3layer = keras.layers.Conv2D(
4    filters=32,
5    kernel_size=(3, 3),
6    padding="valid"
7)

With valid padding, the output becomes smaller than the input when the kernel does not fit at the border without dropping positions.

Use padding="same" to Preserve Spatial Size More Closely

"same" pads the input so the output spatial dimensions stay aligned with the input more closely, especially when stride is 1.

1from tensorflow import keras
2
3layer = keras.layers.Conv2D(
4    filters=32,
5    kernel_size=(3, 3),
6    padding="same"
7)

This is a common choice when you want to stack several convolutions without shrinking the feature map at every layer.

Custom Pixel Padding Uses a Separate Layer

Conv2D itself does not let you pass an arbitrary tuple like "pad two pixels on the left and one on the right" through the padding argument. For that, use an explicit padding layer.

1from tensorflow import keras
2
3model = keras.Sequential([
4    keras.layers.ZeroPadding2D(padding=((1, 1), (2, 2))),
5    keras.layers.Conv2D(32, (3, 3), padding="valid")
6])

This is the standard way to get exact manual padding behavior in Keras.

Padding Affects Output Shape and Edge Behavior

Padding is not only a shape-setting convenience. It changes which border information the model can use and how quickly feature maps shrink across the network.

That means the padding choice is part of architecture design, not only syntax. A model using many valid convolutions may shrink feature maps aggressively. A model using same padding preserves more spatial alignment across layers.

That architectural effect becomes easier to notice in deeper CNN stacks, where repeated border shrinkage compounds across many layers.

Stride Still Matters

"same" does not mean output size is literally identical in every possible case. Stride interacts with the formula too.

That is why it helps to think of same padding as "preserve alignment as designed by the framework" rather than as a magical guarantee of unchanged shape under every parameter combination.

Keep the Intention Clear in Model Code

If the architecture needs exact explicit border handling, say so with a padding layer. If the architecture only needs standard no-padding or same-padding behavior, use the built-in argument directly.

This keeps the model readable. Future readers should be able to see whether the padding is a generic architectural choice or a carefully tuned border design.

Readable model code pays off when you revisit training behavior months later.

It also makes debugging tensor shapes much less painful.

That alone saves time on real projects.

Small shape decisions propagate through an entire CNN.

That compounds quickly in deep models.

Common Pitfalls

• Expecting Conv2D(padding=...) to accept arbitrary custom pixel widths directly.
• Confusing "same" with "always identical output shape no matter what".
• Forgetting that padding changes more than just tensor dimensions.
• Using valid padding repeatedly and then being surprised by rapid spatial shrinkage.
• Hiding important custom border behavior instead of making it explicit with a padding layer.

Summary

• In Conv2D, the common built-in choices are "valid" and "same".
• Use "valid" for no padding.
• Use "same" when you want standard padding behavior that preserves spatial alignment more closely.
• Use ZeroPadding2D for explicit custom border widths.
• Treat padding as an architectural choice, not just a syntax detail.