numpix

An almost dependency free (as far as I'm concerned, numpy is part of the stdlib of Python haha) array visualizer for the terminal. Inspect numpy, PyTorch, and JAX arrays as colored pixel grids — right in your shell. Think treescope but without the jupyter part.

Features

• Works with numpy, PyTorch, and JAX — pass any array-like, no conversion needed
• 1D, 2D, and 3D arrays — 1D arrays render as a row, 3D arrays render slices side-by-side
• Kitty graphics protocol — crisp, high-resolution rendering in Kitty, Ghostty, and WezTerm; graceful Unicode fallback everywhere else
• Stats header — shape, min, max, mean, and variance printed above every visualization
• Multiple colormaps — cividis, magma, inferno, plasma, hot, coolwarm, grey
• Large array truncation — big arrays are intelligently truncated with a grey separator so you always see both ends
• Zero dependencies — pure Python stdlib + numpy only

Examples

Fallback for terminals that don't support the kitty protocol:

Installation

pip install numpix

Usage

from numpix import pix
import numpy as np

# 2D array
pix(np.random.rand(28, 28))

# 1D array — rendered as a single row
pix(np.sin(np.linspace(0, 2 * np.pi, 100)))

# 3D array — slices shown side by side
batch = np.stack([
    np.random.rand(28, 28),
    np.eye(28),
    np.outer(np.sin(np.linspace(0, np.pi, 28)), np.ones(28)),
])
pix(batch)

Works with PyTorch and JAX tensors too — no .numpy() needed:

import torch, jax.numpy as jnp

pix(torch.randn(3, 28, 28))
pix(jnp.array(batch))

Multiple arrays

Pass multiple arrays to pix() to visualize them one after another. Each array gets its own color range by default:

a = np.sin(np.linspace(0, 4 * np.pi, 100)).reshape(10, 10)
b = np.random.rand(10, 10) * 100
pix(a, b)

Use shared_range=True to normalize all arrays to the same color scale — useful for seeing how values evolve across arrays:

pix(a, b, shared_range=True)

Arrays with inf values

pix handles -inf and inf gracefully — -inf maps to the darkest color, inf to the brightest:

masked = np.random.randn(8, 8)
masked[np.tril_indices(8, -1)] = -float("inf")
pix(masked)

API

numpix.pix(
    *arrays,
    max_show: int = 40,
    color_scheme: str = "cividis",
    max_slices: int = 3,
    layout: Literal["horizontal", "vertical"] = "horizontal",
    use_kitty_protocol: bool = True,
    shared_range: bool = False,
    scale: int | Literal["auto"] = "auto",
)

Terminal support

numpix auto-detects support at import time — no configuration required.

License

MIT