When Should Plt Subplots Figsize Use Inches Versus Pixels?

Oh man, fiddling with figsize and dpi in plt.subplots is one of those tiny pleasures that makes a figure go from meh to crisp. At the core it's simple math: figsize is in inches, dpi is dots (pixels) per inch, so pixel dimensions = (width_inches * dpi, height_inches * dpi). For example, fig, axes = plt.subplots(2, 2, figsize=(6, 4), dpi=100) results in a 600×400 pixel canvas. That total canvas is then divided among the subplots (plus margins), so each axes’ drawable area scales with those numbers. I often do the math mentally when I want a specific pixel size for a web thumbnail or a poster panel. Where it gets juicy is how text, line widths, and rasterization behave. Font sizes are typically in points (1 point = 1/72 inch), so their physical size on the figure stays consistent with figsize and dpi; bumping dpi increases pixel density, making text and lines crisper without changing their physical inch size. But saving is another twist: plt.savefig has its own dpi argument that overrides fig.dpi — handy if I make a quick onscreen 100 dpi fig but need a 300 dpi export for printing. Also, vector formats like 'pdf' or 'svg' don't rasterize curves at a given dpi, so they stay sharp when scaling; however embedded raster images or artists that are rasterized will still depend on dpi. Practical tips I use: set figsize to control layout and spacing (how many subplots comfortably fit), use dpi to control resolution for output, and prefer vector formats for publication. If you're stacking many subplots, tweak figsize first, then adjust dpi if you need more pixel detail. I usually test with a small export at different dpi values until the labels and tick marks look right — it's almost satisfying, like fine-tuning a synth patch.

Oh, this is one of those tiny plotting details that trips people up at first, but once you see how Matplotlib behaves it starts to make sense. When you call plt.subplots(figsize=(w, h)) you are setting the initial size of the Figure in inches. On interactive backends (like Qt5Agg, TkAgg, etc.) the figure lives inside a resizable GUI window, and when that window changes size the canvas pixel dimensions change too. Because Matplotlib places axes and subplots using normalized figure coordinates, the axes themselves scale with the window, so visually the subplots do resize as the window is resized. That said, there are caveats. figsize is a stored property for the figure and reflects the current figure size in inches; it was set initially but can update if the window is resized (since inches = pixels / dpi). However, spacing between subplots (margins, padding) is not always recomputed automatically in the way you might expect. If you need spacing recalculated on resize, use constrained_layout=True when creating the figure or call fig.tight_layout() after a resize. For full control you can register a resize callback with fig.canvas.mpl_connect('resize_event', callback) and inside the callback call fig.set_size_inches(...) or fig.tight_layout() and then fig.canvas.draw_idle(). In short: yes, interactive backends will visually resize your subplots when the window changes, but for consistent layout behavior you may want constrained_layout, tight_layout, or a resize handler that updates spacing and forces a redraw.

Oh, matplotlib sizing is one of those little puzzles I tinker with whenever a figure looks either cramped or ridiculously spacious. Figsize in plt.subplots is simply the canvas size in inches — a tuple like (width, height). That number doesn't directly set the gap between axes in absolute terms, but it strongly affects how those gaps look because it changes the total real estate each subplot gets. Practically, spacing is controlled by a few things: wspace/hspace (fractions of average axis size), fig.subplots_adjust(left, right, top, bottom, wspace, hspace) (normalized coordinates), and auto-layout helpers like tight_layout() and constrained_layout=True. For instance, wspace is a fraction of the average axis width; if you make figsize bigger, that same fraction becomes a larger physical distance (more inches/pixels), so subplots appear further apart. DPI multiplies inches to pixels, so a (6,4) figsize at 100 DPI is 600x400 pixels — larger DPI increases resolution but not the inch spacing. I like practical snippets: fig, axs = plt.subplots(2,2, figsize=(8,6), gridspec_kw={'wspace':0.25,'hspace':0.35}); or fig.subplots_adjust(wspace=0.2, hspace=0.3). If labels or legends overlap, try fig.set_constrained_layout(True) or fig.tight_layout(). Also consider gridspec_kw with width_ratios/height_ratios or using GridSpec directly for fine control. Bottom line: figsize sets the stage; subplots_adjust, wspace/hspace, and layout engines direct the actors. Play with the DPI and constrained_layout until everything breathes the way you want — I often tweak it when saving figures for papers versus slides.

Oh, this plotting little puzzle is one of my favorites to tinker with! If you want plt.subplots(figsize=...) to preserve an aspect ratio, the trick is knowing that figsize controls the overall figure inches, while axes have their own box and data aspect settings. For simple cases I like to set the axes box aspect so the axes themselves keep the width:height ratio I want: ax.set_box_aspect(h/w) (requires Matplotlib 3.3+). That makes the axes rectangle scale correctly no matter how the figure is resized. A practical pattern I use a lot: compute the total figure size from the number of columns and rows and your desired per-axis aspect. For example, if each subplot should be 4:3 (width:height) and you have 3 cols and 2 rows, pick a base width (say 3 inches per subplot) and set figsize=(3*3, 3*3*(3/4)) or more simply derive height = width * (rows/cols) * (desired_height/desired_width). Then set constrained_layout=True or tight_layout() so Matplotlib honors margins and suptitles without clipping. Example sketch: fig, axes = plt.subplots(2, 3, figsize=(9, 6), constrained_layout=True) for ax in axes.flat: ax.set_box_aspect(3/4) # keeps each axis box at 3:4 (h/w) so the images look right If you must preserve data units (one x unit equals one y unit), use ax.set_aspect('equal', adjustable='box') instead. For images, imshow(..., aspect='equal') or set extent so axes scaling is consistent. Also watch out: colorbars, legends, and titles change free space, so either reserve space with GridSpec or use set_box_aspect so the axes ignore figure decorations when keeping shape. I like this approach because it’s deterministic — you get square-ish or fixed-ratio panels without manual fiddling.

Funny little quirk, right? I used to be bamboozled by this until I dug in: figsize is measured in inches, dpi is dots-per-inch, and tight_layout is purely a layout algorithm — they all live in the same universe but play different roles. Figsize = (width, height) in inches. DPI = pixels per inch. So the expected pixel dimensions are figsize * dpi. tight_layout, though, doesn't change those inches or DPI. What tight_layout actually does is adjust subplot parameters (left/right/top/bottom, spacing) so axes, labels and titles fit inside the figure canvas. It can shrink the effective axes area or add space around them, but it doesn't rewrite fig.get_size_inches() or fig.dpi. Where people see 'ignored DPI' is usually later: when you call savefig with bbox_inches='tight' or display in a notebook, Matplotlib crops or rescales the image bounding box, and savefig has its own dpi parameter that can override or interact with figure.dpi. Practical checklist that helped me: check fig.get_size_inches() and fig.dpi before and after tight_layout; call fig.canvas.draw() to ensure layouts are computed; if saving, use savefig(dpi=...) explicitly and be careful with bbox_inches='tight' because it crops and may change pixel dimensions; if you’re in a high-DPI (retina) display, the display backend can scale the figure differently. If you want absolute control, set figsize and dpi, call fig.set_dpi(...), avoid bbox_inches='tight' or compute the bounding box yourself, or use plt.subplots_adjust to lock margins. Once I started thinking in inches + dpi + cropping as three separate steps, things clicked.

Funny little detail that trips people up: matplotlib's figsize is measured in inches. I say this like someone chatting over coffee with a sketchpad of plots — figsize=(6,4) means 6 inches wide and 4 inches tall, not pixels, not centimeters. The reason that matters is DPI (dots per inch) — matplotlib uses the figure's DPI to convert those inches into pixels. By default, modern matplotlib sets figsize to (6.4, 4.8) inches and dpi to 100, so a default figure ends up being 640×480 pixels when rendered or saved (6.4*100 by 4.8*100). In practice I often treat figsize like the physical size of a poster: if I need a poster for a talk or a high-res image for a paper, I pick bigger inches and/or bump dpi when saving. For example, figsize=(8,6) with dpi=200 gives 1600×1200 pixels. You can set dpi in plt.figure(..., dpi=...) or override it at save time with savefig(..., dpi=300). If you want to inspect or change a created figure you can use fig.get_size_inches() and fig.set_size_inches(w,h). Tiny pro tip from my late-night tinkering: if you prefer metric, multiply inches by 2.54 to get centimeters. When embedding in notebooks some backends or frontends scale images visually, so pixel counts might feel off — but mathematically, figsize is always inches and the DPI does the conversion. I find thinking about inches helps when preparing figures for print or slides, and it makes resizing less mysterious.

Totally—yes, but not by using figsize on each subplot directly. Figsize controls the overall figure canvas size (the whole window or saved image), not individual axes. If you want subplots with different widths or heights, I usually reach for GridSpec-style tools or explicit axes placement. In practice I do something like this: plt.subplots(figsize=(10,6), gridspec_kw={'width_ratios':[3,1]}) to make the left plot three times wider than the right one. For more control I create a GridSpec: gs = fig.add_gridspec(2,2, width_ratios=[2,1], height_ratios=[1,2]) and then use fig.add_subplot(gs[0,0]) and so on. If I need pixel-precise placement, fig.add_axes([left, bottom, width, height]) with normalized coordinates (0–1) is my go-to. There are also helpers like make_axes_locatable or inset_axes if you want a small inset plot or colorbar region attached to a main axis. A couple of practical tips from projects where I fussed over layouts: use gridspec_kw with plt.subplots for quick proportional layouts, try constrained_layout=True or fig.tight_layout() to avoid overlaps, and remember that aspect and axis labels can change perceived sizes. For interactive tweaking, I often use notebook sliders or tiny scripts that print axis.get_position() so I can fine-tune left/right values. Happy plotting — once you get the grid ratios right, it feels like arranging panels in a comic strip, which always makes me smile.

Occasionally I tweak a figure's size and the legend seems to shift like it has a mind of its own — that's normal, and here's why it happens. When you call plt.subplots(figsize=(w,h)) you're changing the figure's pixel dimensions (and thus the axes' size and positions). Legends are positioned relative to axes or the whole figure depending on how you create them: ax.legend() uses the axes coordinate system by default (so 'upper right' is inside the axes), while fig.legend() anchors to the figure. Changing figsize alters the underlying coordinate-to-pixel mapping, so a legend that was fine in a tiny figure can look crowded or appear to sit in a different spot in a larger one. Beyond that, layout managers like tight_layout() and constrained_layout=True will try to rearrange axes and decorations (including legends) to avoid overlap. If you use bbox_to_anchor, its coordinates are interpreted in the transform you choose — often ax.transAxes or fig.transFigure — and that decides whether the anchor scales with the axes or the whole figure. DPI matters too: a bigger figsize with the same DPI increases pixel space, which can make elements seem farther apart. In practice I fix placement explicitly: use ax.legend(loc='center left', bbox_to_anchor=(1,0.5)) with bbox_transform=ax.transAxes to pin it relative to the axes, or use fig.legend(...) with fig.transFigure when I want a shared legend that follows figure size. If things get clipped on save, supply bbox_inches='tight' or add bbox_extra_artists to savefig, or tweak subplots_adjust. Little experiments with loc, bbox_to_anchor and transforms usually get the behavior I want.