tephpy add_logo — design specification#

Estimated reading time: 23 minutes

Living document. This specification is maintained alongside the code, not archived behind it. src/tephpy/plotting/logo.py cites it by section — logo spec §3.5 and the like — so these sections are the reasoning behind what the code does, and where the two ever diverge it is the specification that gets corrected. Read it as current.

  • Date: 2026-08-01 (originated; maintained since)

  • Status: living design specification, implemented in #71

  • Scope: one new public function, tephpy.plotting.add_logo, plus six bundled PNG masters

  • Parent spec: 2026-07-22-tephpy-design.md — this extends spec §3.2 plotting with a branding artist and inherits its error-handling (spec §6), testing (spec §7) and engineering-standards (spec §8) rules unchanged

  • Prior art: MetPy’s add_metpy_logo

  • Brand assets: docs/src/_static/brand/assets/logo-bundle.zip, catalogued by docs/src/_static/brand/assets/README.md (PR #69)

1. Purpose#

Give users a one-call way to brand a figure:

tephpy.plotting.add_logo(ax)

MetPy proved the demand for this and the shape of the API. It also shows the two things not to copy. add_metpy_logo uses Figure.figimage, which places the image in device pixels — so the same call yields a logo a third the relative size at 300 dpi that it had at 100 dpi, and there is no way to ask for “half an inch tall”. And it offers a single dark-on-light raster, so on a dark figure the mark disappears into the background.

tephpy has richer inputs to work with: three logo forms (icon, lockup, stacked) each in a light-background and a dark-background variant, all published in the brand bundle. The design below spends those inputs on a dpi-independent, theme-aware placement built from AnnotationBbox, whose positioning vocabulary users already know from legend.

2. Decisions#

Decision

Choice

Rationale

Artist

AnnotationBbox(OffsetImage(...))

Sizes in points, so a request in inches renders identically at every dpi. figimage (MetPy’s choice) sizes in device pixels and cannot honour an inch request

Sizing units

Height in inches; width follows the master’s aspect

One number per form; the caller never reasons about pixels or aspect ratios

Masters shipped

The largest of each form (icon 512, lockup 716, stacked 512), light and dark

128.3 KiB for six files. One master per (form, variant) downscales to every size; shipping the smaller rungs too would add weight and a second thing to keep in sync

Positioning vocabulary

legend’s loc= — ten strings or an (x, y) tuple

Zero new vocabulary. 'best' is deliberately excluded: it promises collision detection this artist does not do

What the position is relative to

The target — a Figure anchors in figure fraction, an Axes in axes fraction

Same rule as legend: the thing you call it on defines the box. No separate relative_to= kwarg

Default form

"lockup"

It carries the wordmark at the smallest height of the three (§3.3), so the brand is legible in the least plot area. A bare icon is meaningless to anyone who does not already know the brand

Theme selection

theme="auto" from target facecolor luminance

Correct without being asked on both the default white figure and a dark_background style. "light"/"dark" name the background, matching the asset filenames

Asset packaging

Copies under src/tephpy/plotting/_static/, declared in [tool.setuptools.package-data] and MANIFEST.in

The function must work from a wheel with no docs tree and no network. Measured: while the files are git-tracked, include-package-data (on by default) plus the setuptools_scm file finder already ships them — the declaration is belt-and-braces that survives losing either, not the load-bearing mechanism, and the spec does not pretend otherwise

Asset drift

A test hashes the six copies against logo-bundle.zip

The zip is the source of truth; the copies are derived. Without the guard they diverge silently on the next rebrand

Return value

The AnnotationBbox

The caller can restyle or .remove() it. Matches matplotlib’s artist-returning convention

Tinted mono variant (color=)

Deferred (#72)

Needs the mono SVGs rasterised offline; the light masters cannot serve as an alpha mask because they are three-tone with substantial white knockout (§8)

3. Architecture#

src/tephpy/plotting/
├── logo.py                     # add_logo + the three lookup tables
└── _static/
    ├── icon-512-light.png      # 512 x 512
    ├── icon-512-dark.png
    ├── lockup-716-light.png    # 716 x 256
    ├── lockup-716-dark.png
    ├── stacked-512-light.png   # 512 x 720
    └── stacked-512-dark.png

One new module, one new package-data directory, one new public name. logo.py imports nothing from tephpy beyond _constants; it does not touch transforms, axes or isopleths, and nothing imports it back. tephpy.plotting.__init__ grows from tephpy.plotting.logo import add_logo and the matching __all__ entry.

3.1 Public API#

def add_logo(
    target: Figure | Axes | None = None,
    *,
    form: str = "lockup",
    size: str | float = "small",
    theme: str = "auto",
    loc: str | tuple[float, float] = "lower left",
    pad: float | None = None,
    zorder: float | None = None,
    **kwargs: Any,
) -> AnnotationBbox:

pad and zorder take None sentinels rather than literals because their values are numeric conventions, and the rule in spec §3.5 is that nothing numeric is hard-coded at point of use — a signature default is a point of use. They resolve to LOGO_PAD (6.0) and LOGO_ZORDER (100.0) from _constants. The string defaults stay literal: they are API vocabulary, not conventions, and reading them off the signature is worth more than routing them through a constant.

Parameter

Accepts

Meaning

target

Figure, Axes, None

What to brand, and what the position is relative to. None → the current figure (plt.gcf())

form

"lockup", "icon", "stacked"

Which logo

size

"small", "large", or a float

Preset (§3.3) or an explicit height in inches

theme

"auto", "light", "dark"

Which variant; the name is the background the logo is drawn on

loc

ten legend strings, or (x, y)

Where (§3.4)

pad

float, None

Points between the logo and the target edge; None takes LOGO_PAD. Ignored when loc is a tuple

zorder

float, None

Draw order; None takes LOGO_ZORDER, which sits above lines (2), text (3) and legends (5)

**kwargs

alpha, interpolation, resample, filternorm, filterrad

Forwarded to OffsetImage

target=None resolves through plt.gcf(), so add_logo() on its own brands the current figure — the MetPy-equivalent zero-argument call.

The artist is attached with fig.add_artist(ab) for a Figure target and ax.add_artist(ab) for an Axes target; the latter is what makes xycoords="axes fraction" resolvable.

3.2 Bundled assets and packaging#

The six masters are byte-for-byte copies of bundle/png/ members of docs/src/_static/brand/assets/logo-bundle.zip. They are resolved through one table so a third variant can be added without touching the call site:

_MASTERS = {
    ("icon", "light"): "icon-512-light.png",
    ("icon", "dark"): "icon-512-dark.png",
    ("lockup", "light"): "lockup-716-light.png",
    ("lockup", "dark"): "lockup-716-dark.png",
    ("stacked", "light"): "stacked-512-light.png",
    ("stacked", "dark"): "stacked-512-dark.png",
}

Loading is lazy, via importlib.resources.files("tephpy.plotting") / "_static" / name, read with matplotlib.image.imread. Nothing is read at import time, so import tephpy costs nothing extra and the module honours the existing import-discipline test at tests/plotting/test_isopleths.py:186.

Packaging is declared explicitly:

[tool.setuptools.package-data]
tephpy = ["py.typed", "plotting/_static/*.png"]

py.typed is listed alongside because it currently ships with no declaration at all, relying on the setuptools_scm file finder — the same implicit mechanism this table replaces for the PNGs.

3.3 Sizing#

size is a height in inches. Width follows from the master’s aspect ratio, so a form is never distorted and the caller supplies one number.

form

Master

Aspect (w/h)

"small"

"large"

Rendered at "small"

lockup

716 × 256

2.797

0.30 in

0.55 in

0.84 × 0.30 in

stacked

512 × 720

0.711

0.70 in

1.15 in

0.50 × 0.70 in

icon

512 × 512

1.000

0.40 in

0.70 in

0.40 × 0.40 in

The presets differ per form because the forms place the wordmark differently. Measured on the light masters, the wordmark occupies 44.1% of the lockup’s height but only 17.8% of the stacked form’s. Setting the "small" presets so the wordmark clears ~12 px at dpi 100 gives 13.2 px for the lockup at 0.30 in and 12.4 px for the stacked form at 0.70 in — the stacked form needs 2.3× the height for the same legibility. A single shared preset would either shrink the stacked wordmark to mush or make the lockup gratuitously large. "large" is 1.6–1.8× "small" in each case. The icon carries no wordmark and so has no legibility floor to meet; its presets sit between the other two forms’.

3.4 Placement#

loc takes matplotlib’s legend vocabulary minus 'best': 'upper right', 'upper left', 'lower left', 'lower right', 'right', 'center left', 'center right', 'lower center', 'upper center', 'center'. As in legend, 'right' is a synonym for 'center right', so the ten strings name nine positions. 'best' is rejected with a message saying so — this artist does no collision detection and silently aliasing it to a corner would be a lie.

One table drives all of them: each string maps to an anchor in target-fraction coordinates, a box_alignment, and the sign of the padding offset.

loc

anchor xy

box_alignment

offset (points)

lower left

(0, 0)

(0, 0)

(+pad, +pad)

lower right

(1, 0)

(1, 0)

(−pad, +pad)

upper left

(0, 1)

(0, 1)

(+pad, −pad)

upper right

(1, 1)

(1, 1)

(−pad, −pad)

lower center

(0.5, 0)

(0.5, 0)

(0, +pad)

upper center

(0.5, 1)

(0.5, 1)

(0, −pad)

center left

(0, 0.5)

(0, 0.5)

(+pad, 0)

center right, right

(1, 0.5)

(1, 0.5)

(−pad, 0)

center

(0.5, 0.5)

(0.5, 0.5)

(0, 0)

Anchoring the box’s own corner to the matching target corner is what keeps the gap constant: because box_alignment tracks the anchor, the offset is a pure inward push and never depends on the logo’s rendered size.

A tuple loc=(x, y) places the logo’s lower-left corner at that point in target-fraction coordinates — the same meaning legend(loc=(x, y)) has — via box_alignment=(0, 0) and a zero offset. pad is ignored in this case, and that is documented on the parameter rather than raised as an error: the caller who gave exact coordinates has already said where they want it.

pad defaults to 6.0 points, a shade wider than legend’s borderaxespad of 5.0 pt (0.5 font-size units at the 10 pt default font).

3.5 Theme resolution#

theme="light" and theme="dark" name the background the logo will sit on, which is also how the asset filenames are named, so there is one vocabulary and no inversion to remember.

theme="auto" reads the target’s facecolor and picks by Rec. 709 luma over the gamma-encoded sRGB channels (0.2126 R + 0.7152 G + 0.0722 B), choosing dark below 0.5 and light at or above it. This is luma, not relative luminance: the latter applies the same weights but linearises each channel first, which scores mid grey #808080 at 0.216 instead of 0.502 and so pulls the crossover well into the light half of the range. Weighting the encoded values keeps the threshold where a reader would put it by eye, which is all it has to do to choose between two artwork files.

An Axes target is read from ax.get_facecolor() and a Figure from fig.get_facecolor(). Because either may be translucent, they are alpha-composited back to front — an assumed white page, then the figure, then the axes — and the luma is measured on the result, so the test matches what shows through rather than what a layer’s own channels say. Judging a layer alone would score 10% black over a white figure at 0.0 and pick the dark mark for a background the reader sees as near-white. The two ends of that range are the behaviour named above: a fully transparent axes — facecolor="none" — contributes nothing and so falls through to the figure, and if the figure is also transparent the assumed white page carries the answer, light. A fully opaque layer hides everything under it exactly as before.

The compositing itself is not add_logo’s: it lives in plotting/_theme.canvas_rgb, because the isopleth label boxes ask the same question and take the answer unthresholded to tint themselves (spec §3.2). Sharing it is what stops the mark and the diagram disagreeing about what colour the background is — the state #173 left the diagram in, where auto correctly drew the dark mark onto a figure whose labels were still boxed for a white page. It is asked about the axes and walks out to the root figure itself, which matters to the label boxes rather than to add_logo: a SubFigure is transparent by default, so an axes in one is over the root through two invisible sheets, and §5’s refusal of a SubFigure target means the walk from an axes add_logo accepted is always the single hop to its figure.

Documented limitation: savefig(transparent=True) does not change any facecolor; it overrides alpha at draw time. auto therefore still sees white and picks light, which is the right answer for a figure destined for a white page and the wrong one for a dark page. Callers in that position pass theme= explicitly. Closing this properly is the deferred color= work (§8).

3.6 Rendering#

zoom = size_inches * 72 / master_height_px

AnnotationBbox(
    OffsetImage(arr, zoom=zoom, **image_kwargs),
    xy=anchor,
    xycoords="figure fraction",      # or "axes fraction"
    xybox=offset,
    boxcoords="offset points",
    box_alignment=alignment,
    frameon=False,
    pad=0.0,
    zorder=zorder,
    annotation_clip=False,
)

OffsetImage’s zoom is points-per-pixel, so size_inches * 72 / master_height_px renders the master at exactly the requested height at any dpi. Verified at dpi 100/300/600: a 0.30 in lockup measures 0.8391 × 0.3000 in every time, with the corner gap exactly 6.00 pt.

pad=0.0 is mandatory, not cosmetic. AnnotationBbox’s default pad=0.4 is in font-size units — 4 pt per side at the 10 pt default — which inflates the rendered box by a constant 0.111 in regardless of dpi or requested size. Left at the default, a "small" lockup measures 0.411 in instead of 0.300 in and a "large" one 0.661 in instead of 0.550 in. This is the single most easily reintroduced bug in the module and §6 pins it with a regression test.

4. Canonical usage#

import matplotlib.pyplot as plt
import tephpy
from tephpy.plotting import add_logo

fig, ax = plt.subplots(subplot_kw={"projection": "tephigram"})
ax.plot(sounding.pressure, sounding.temperature)

add_logo(ax)                                        # lockup, small, auto theme, lower left
add_logo(fig, form="icon", loc="upper right")       # figure corner, not the plot's
add_logo(ax, size="large", loc=(0.35, 0.02))        # exact placement, pad ignored
add_logo(ax, theme="dark", alpha=0.6)               # explicit variant, watermark weight

5. Error handling#

The plotting layer raises builtin exceptions; TephpyError and its subclasses are for user-correctable data input (parent spec §6) and are not used here.

  • Unknown keyword → TypeError whose message contains "unknown option", matching the convention asserted at tests/plotting/test_isopleths.py:363. **kwargs is validated against the allow-list in §3.1 before forwarding, because matplotlib’s own failure for a bad OffsetImage kwarg is an AttributeError reading BboxImage.set() got an unexpected keyword argument ... — the wrong type and a message that names an artist the caller never mentioned.

  • Unknown form, theme, or loc string → ValueError naming the valid set. loc="best" says explicitly that it is unsupported because no collision detection is performed.

  • A size that is neither a preset name nor a positive finite float → ValueError.

  • A loc that is neither a string nor a two-element sequence of floats → TypeError; a two-element sequence holding a non-finite value → ValueError. Coordinates outside [0, 1] are allowed — they place the logo outside the target box, which annotation_clip=False renders and which legend permits for the same reason.

  • A target that is neither a Figure nor an Axes → TypeError. An Axes whose .figure is a SubFigure also raises TypeError: SubFigure is out of scope (§8), and saying so beats returning something typed Figure that is not one.

6. Testing#

Tests live in tests/plotting/test_logo.py, mirroring the source layout.

  • Tables: every _LOC entry, every (form, size) preset, and every theme resolution is asserted, including 'right' resolving identically to 'center right'.

  • Sizing: the rendered AnnotationBbox window extent equals the requested height in inches, asserted at dpi 100, 300 and 600 — the property MetPy’s figimage approach cannot hold.

  • pad=0.0 regression: an explicit test pinning the exact rendered inches, which fails by a constant 0.111 in if the pad=0.0 argument is ever dropped.

  • Placement: corner gaps measure pad points exactly, across dpi and figsize; a tuple loc puts the lower-left corner where asked and ignores pad.

  • Target semantics: an Axes target anchors to the axes box and a Figure target to the figure box, distinguishable because the two boxes differ.

  • Asset drift: the six shipped PNGs hash equal to their logo-bundle.zip counterparts.

  • Packaging: the six PNGs are present in a built wheel — the failure mode a source-tree test cannot see. It guards the outcome, not the mechanism: the mutation that fails it is removing an asset from the source tree, not deleting the package-data line, which include-package-data would cover for.

  • Errors: each case in §5.

  • Import discipline: logo.py reads no asset at import time.

  • Image baseline: one pytest-mpl comparison, per parent spec §7.

7. Documentation#

  • A how-to under docs/src/howtos/, added to that toctree, covering the default call, the three forms, dark figures, and exact placement.

  • add_logo joins the API reference automatically through autoapi; its numpydoc docstring documents every parameter, the return, and each raise, per the repo’s numpydoc-validation hook.

  • docs/src/_static/brand/assets/README.md gains a line recording that six bundle members are also shipped inside the wheel at tephpy/plotting/_static/, so a future rebrand knows to update both.

8. Scope#

In scope: everything in §3.

Deferred (#72) — a color= kwarg tinting a monochrome mark to an arbitrary colour. This is the honest fix for the transparent-background case in §3.5 and for figures whose background is neither light nor dark. It is deferred because it needs lockup-tiera-mono.svg and its siblings rasterised offline into a third variant — tephpy cannot rasterise SVG at runtime (Pillow does not, and the bundle’s own generators emit SVG rather than consume it). The light masters cannot substitute as an alpha mask: measured, they are three-tone with substantial white knockout (19.2% of icon-512-light.png is pure white), so flattening them by alpha collapses the mark.

Rejected (2026-08-01) — explicitly not in scope: collision detection (loc="best"), animation, SubFigure targets, per-artist logo placement on subplots other than through repeated calls, a tephpy.config section (this work adds exactly one public name), and any change to the published brand assets. Each is a deliberate omission rather than an unbuilt intention, so none carries an issue (docs spec §3.5).

9. References#