matplotlib styles mess with placement

[skwde]

I also asked this on stackoverflow (https://stackoverflow.com/questions/79303139/maptloblib-adjust-text-and-errorbar).
Due to a comment I tried again and found that there is a problem occuring only when matplotlib styles are used.

See following example:

import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import scienceplots  # type: ignore
from adjustText import adjust_text

plt.style.use(["science", "notebook"])

fig, ax = plt.subplots()

x = np.linspace(0, 1, 50) * np.pi
y = np.sin(x)

errbar_container = ax.errorbar(x, y, xerr=0.25, yerr=0.05)

texts = []
for idx in range(0, len(x),2):
    texts.append(
        ax.annotate(
            idx,
            xy=(x[idx], y[idx]),
            size="large",
            zorder=100,
        )
    )

adjust_text(
    texts,
    x=x,
    y=y,
    expand=(2, 2),
    objects=[
        *[mpl.transforms.Bbox(seg) for seg in errbar_container.lines[-1][0].get_segments()],
        *[mpl.transforms.Bbox(seg) for seg in errbar_container.lines[-1][1].get_segments()],
    ],
    arrowprops=dict(arrowstyle="->", color="red"),
)

plt.show()

which produces following plot:

Note how 0, 2 and 4 plotted on the line.

Moreover it seems that option like autoalign, only_move are ignored.

Without plt.style.use(["science", "notebook"]) everything looks fine.

So what am I missing here?

[BBC-Esq]

Looking at your issue, there are a few things going on. The scienceplots "science" style typically enables LaTeX rendering, which can affect how text bounding boxes are calculated. But beyond that, I see some issues in the code itself:

1. Bbox coordinate system mismatch

When you pass Bbox objects to the objects parameter:

objects=[
    *[mpl.transforms.Bbox(seg) for seg in errbar_container.lines[-1][0].get_segments()],
    *[mpl.transforms.Bbox(seg) for seg in errbar_container.lines[-1][1].get_segments()],
],

These segment coordinates are in data coordinates, but the get_bboxes function returns them unchanged when it detects they're already Bbox objects:

if all([isinstance(obj, matplotlib.transforms.BboxBase) for obj in objs]):
    return objs  # ❌ Returned as-is, no transform to display coordinates!

Meanwhile, text bounding boxes are in display coordinates. This mismatch means the collision detection between texts and your errorbar objects doesn't work correctly.

2. LaTeX rendering timing

The "science" style uses LaTeX for text rendering. LaTeX text bounding boxes may not be accurately computed until the figure is fully rendered. The force_draw() call might not be sufficient for LaTeX text.

---

Suggested fix

Transform your Bbox objects to display coordinates before passing them:

transform = ax.transData  # Data -> Display coordinate transform

objects_transformed = []
for seg in errbar_container.lines[-1][0].get_segments():
    bbox_data = mpl.transforms.Bbox(seg)
    bbox_display = bbox_data.transformed(transform)
    objects_transformed.append(bbox_display)
for seg in errbar_container.lines[-1][1].get_segments():
    bbox_data = mpl.transforms.Bbox(seg)
    bbox_display = bbox_data.transformed(transform)
    objects_transformed.append(bbox_display)

adjust_text(
    texts,
    x=x,
    y=y,
    expand=(2, 2),
    objects=objects_transformed,
    arrowprops=dict(arrowstyle="->", color="red"),
)

Or, you might also try forcing a full draw before adjust_text:

fig.canvas.draw()
adjust_text(...)

This can help ensure LaTeX text bounding boxes are properly calculated before the adjustment algorithm runs.