dragonsci

GPU-accelerated scientific visualization for Python, written in Rust.

Renders large point clouds (250 k+ points) at interactive frame rates by using wgpu (Vulkan / Metal / DirectX 12) via PyO3. The widget embeds directly in any Tkinter layout — no separate window or subprocess.

Installation

Tkinter (default):

pip install dragonsci

JupyterLab / Jupyter Notebook:

pip install "dragonsci[notebook]"

The [notebook] extra adds jupyter_rfb, ipywidgets, and Pillow.

The release workflow builds pre-built wheels for Python 3.9–3.12 on Windows, macOS, and Linux, plus a source distribution for source installs.

GitHub release publishing

This repo now includes automated PyPI publishing through release.yml. The workflow builds:

• a source distribution
• Linux x86_64 manylinux wheels
• Windows x64 wheels
• macOS x86_64 and arm64 wheels

Publishing uses PyPI Trusted Publishing. Before the first release, configure dragonsci on PyPI to trust this exact GitHub Actions workflow:

1. Repository owner: NKocur
2. Repository name: dragonsci
3. Workflow path: .github/workflows/release.yml

After that, publishing a GitHub Release will build artifacts and upload them to PyPI automatically. You can also run the same workflow manually from GitHub Actions via workflow_dispatch; leave publish unchecked for a build-only dry run, or enable it to upload to PyPI.

Quick start

Tkinter

import tkinter as tk
import numpy as np
from dragonsci import Scatter3D

root = tk.Tk()
widget = Scatter3D(root, width=900, height=700)
widget.pack(fill="both", expand=True)

pts = np.random.default_rng(0).standard_normal((250_000, 3)).astype(np.float32)
widget.set_points(pts, colormap="plasma")

root.mainloop()

JupyterLab / Jupyter Notebook

from dragonsci import scatter3d   # requires dragonsci[notebook]
import numpy as np

w = scatter3d(width=800, height=600)
w.set_points(np.random.randn(50_000, 3).astype("f4"), colormap="viridis")
w  # display in cell

For 2-D scatter plots pass (N, 2) or (N, 3) arrays to JupyterScatter2D:

from dragonsci import JupyterScatter2D
import numpy as np

w = JupyterScatter2D(width=700, height=500)
w.set_points(np.random.randn(20_000, 2).astype("f4"), colormap="plasma")
w

Point picking is available via callback:

def on_hit(result):
    print(result["actor"], result["index"], result["point"])

w.enable_point_picking(on_pick=on_hit)

API

Widget construction

Scatter3D(master, *, width=800, height=600, fps=60, vsync=True, **kwargs)

A tk.Frame subclass. All standard frame keyword arguments are forwarded.

Pre-map behaviour: The following may be called before mainloop() and are queued, then replayed when the renderer initialises on widget map: set_points, add_points, add_lines, add_box, scalar_bar, show_orientation_axes, point_style, show_grid, set_background, set_axes, set_ticks, and the parallel_projection property.

Camera methods (reset_camera, view_*, fit, get_camera, set_camera) and overlay-mutation methods (update_lines, remove_overlay, set_overlay_visibility) are silent no-ops before the widget is mapped.

Point clouds

widget.set_points(positions, *, colors=None, scalars=None, colormap="viridis",
                  point_size=4.0, clim=None, nan_color=(0.3,0.3,0.3),
                  log_scale=False, opacity=1.0)

Replace all point actors with a single point cloud. Line overlays are not affected; call clear() first to reset the whole scene.

handle = widget.add_points(positions, *, ...)   # same kwargs as set_points
widget.update_actor(handle, positions, *, ...)  # replace in-place
widget.set_actor_visibility(handle, visible)
widget.remove_actor(handle)
widget.clear()                                   # remove all actors, overlays, and grid

Rendering style

widget.point_style = "circle"   # "circle" | "square" | "gaussian"
widget.set_ticks(x=None, y=None, z=None)

Camera

widget.reset_camera()
widget.fit(bounds=None)          # (xmin,ymin,zmin,xmax,ymax,zmax) or re-fit data
widget.view_xy(); widget.view_xz(); widget.view_yz(); widget.view_isometric()
widget.parallel_projection       # bool property
state = widget.get_camera()
widget.set_camera(state)

Axis labels, grid and background

widget.set_axes(x="X", y="Y", z="Z")   # axis title labels at grid extents
widget.show_grid(visible=True)          # toggle grid lines + tick labels
widget.set_background(color)            # (r,g,b) tuple or "#RRGGBB" hex string

Scalar bar

widget.scalar_bar(visible=True, *, vmin=0.0, vmax=1.0, log_scale=False,
                  colormap="viridis", title="")

Line overlays

Segments are (N, 6) float32 arrays, each row [x0,y0,z0, x1,y1,z1]. Overlay bounds are included in camera fitting and grid extents.

handle = widget.add_lines(segments, color=(1,1,1))
handle = widget.add_box(bounds, color=(1,1,0))   # (xmin,ymin,zmin,xmax,ymax,zmax)
widget.update_lines(handle, segments, color)
widget.set_overlay_visibility(handle, visible)
widget.remove_overlay(handle)
widget.clear_overlays()
widget.show_orientation_axes(visible=True)

Picking

widget.enable_point_picking(on_pick=None)      # fires <<PointPicked>>
widget.enable_rectangle_picking(on_select=None) # Shift+drag; fires <<SelectionChanged>>
widget.disable_picking()
# After <<PointPicked>>:  widget.picked_point, .picked_index, .picked_actor
# After <<SelectionChanged>>: widget.selected  →  [{"actor":int,"index":int},...]

Animation export

widget.open_gif(path, fps=20, loop=0)
widget.write_frame()
widget.close_gif()
# or all-in-one orbit:
widget.orbit_gif(path, n_frames=60, fps=20, loop=0, elevation=0.3, on_progress=None)

Screenshot

rgba = widget.screenshot()   # (H, W, 4) uint8 ndarray, or None before map
widget.save_png(path)

Linked cameras

from dragonsci import link_cameras, unlink_cameras
link_cameras(w1, w2, w3, ...)
unlink_cameras(w1, w2)

Misc

Scatter3D.colormap_names() → list[str]

Camera controls

Platform notes

Windows

Fully supported. Uses DirectX 12 or Vulkan via wgpu. No extra dependencies.

macOS

Fully supported. Uses Metal via wgpu. No extra dependencies.

Linux

Requires an X11/Xlib display (Wayland is not supported). The widget uses winfo_id() to obtain an XID and opens the display via libX11.so.6. Pure Wayland sessions will fail at renderer initialisation with a warning. Running under XWayland works.

# Confirm X11 is available:
echo $DISPLAY   # should print e.g. :0

Building from source

Requires Rust (stable) and maturin ≥ 1.7.

Development (installs the extension in-place for local imports):

pip install maturin
maturin develop --release
pytest tests/

Distribution wheel — use the helper script, which removes the maturin develop artifact from the source tree before building. Running maturin build directly from a development working tree will fail because the in-place extension collides with the freshly-compiled one:

python scripts/build_wheel.py           # produces target/wheels/*.whl
# or equivalently:
rm -f python/dragonsci/_dragonsci*.{pyd,so,dylib,pdb}
maturin build --release

License

MIT