diff --git a/DESCRIPTION b/DESCRIPTION
index 6c9a986..697140a 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -46,12 +46,18 @@ Imports:
 Suggests:
     BiocStyle,
     chromote,
+    htmlwidgets,
     knitr,
+    magick,
     rmarkdown,
     shinytest2,
     spelling,
     testthat (>= 3.0.0),
+    webshot2,
     withr
+SystemRequirements: Google Chrome or Chromium (only for
+    exportAMRVisualizations(), which drives a headless browser via
+    webshot2/chromote to render the interactive plots to static files)
 VignetteBuilder: knitr
 biocViews:
     Software,
diff --git a/NAMESPACE b/NAMESPACE
index f7c43c8..2f6f3e1 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -1,5 +1,6 @@
 # Generated by roxygen2: do not edit by hand
 
+export(exportAMRVisualizations)
 export(launchAMRDashboard)
 import(shiny)
 importFrom(DT,datatable)
diff --git a/R/export.R b/R/export.R
new file mode 100644
index 0000000..5a2c388
--- /dev/null
+++ b/R/export.R
@@ -0,0 +1,768 @@
+# Headless batch export of every dashboard visualization to static image
+# files. Lets a user render the full set of amRviz figures from the packaged
+# demo data (or their own amRml results) without ever launching the Shiny app.
+
+
+#' Snapshot one htmlwidget to static image file(s)
+#'
+#' Saves the widget to a temp HTML bundle, renders it in a headless Chrome,
+#' then writes each requested format. Raster formats (png/jpg/pdf) all derive
+#' from a single webshot PNG so they share crop and dimensions; svg is
+#' extracted from the rendered DOM via chromote.
+#'
+#' @param widget An htmlwidget (plotly / networkD3), or NULL.
+#' @param path_base Output path without extension; each format is written as
+#'   `path_base.<ext>`.
+#' @param formats Character vector of extensions among png/pdf/jpg/jpeg/svg.
+#' @param width,height Snapshot viewport size in pixels.
+#' @param scale Device-pixel multiplier for the underlying PNG render.
+#' @param trim When TRUE, `magick::image_trim()` crops the PNG before other
+#'   raster formats are derived from it. Meant for figures with known excess
+#'   whitespace (currently the drug-feature network only).
+#' @param delay Seconds to let the widget's JavaScript render before capture.
+#' @param verbose Whether to message on per-format failures.
+#' @return Character vector of files actually written.
+#' @keywords internal
+#' @noRd
+.exportWidgetFile <- function(widget, path_base, formats,
+                              width = 1200, height = 800, scale = 2,
+                              trim = FALSE,
+                              delay = 1.5, verbose = TRUE) {
+  if (is.null(widget)) return(character(0))
+  written <- character(0)
+  raster <- intersect(formats, c("png", "pdf", "jpg", "jpeg"))
+  want_svg <- "svg" %in% formats
+  if (!length(raster) && !want_svg) return(written)
+
+  tmpdir <- tempfile("amrviz_widget_")
+  dir.create(tmpdir)
+  on.exit(unlink(tmpdir, recursive = TRUE), add = TRUE)
+  html <- file.path(tmpdir, "widget.html")
+  # selfcontained = TRUE inlines JS/CSS so the extracted SVG stands alone.
+  saved <- tryCatch(
+    {
+      htmlwidgets::saveWidget(widget, html, selfcontained = TRUE)
+      TRUE
+    },
+    error = function(e) {
+      if (verbose) message("    saveWidget failed: ", conditionMessage(e))
+      FALSE
+    }
+  )
+  if (!saved) return(written)
+
+  if (length(raster)) {
+    tmp_png <- file.path(tmpdir, "render.png")
+    rendered <- tryCatch(
+      {
+        webshot2::webshot(
+          html, tmp_png,
+          vwidth = width, vheight = height,
+          zoom = scale, delay = delay, quiet = TRUE
+        )
+        file.exists(tmp_png) && file.info(tmp_png)$size > 0
+      },
+      error = function(e) {
+        if (verbose) message("    render failed: ", conditionMessage(e))
+        FALSE
+      }
+    )
+    if (rendered) {
+      if (trim) .trim_raster(tmp_png, verbose = verbose)
+      for (ext in raster) {
+        out <- paste0(path_base, ".", ext)
+        ok <- if (ext == "png") {
+          file.copy(tmp_png, out, overwrite = TRUE)
+        } else {
+          .convert_raster(tmp_png, out, ext, verbose = verbose)
+        }
+        if (isTRUE(ok)) written <- c(written, out)
+      }
+    }
+  }
+
+  if (want_svg) {
+    out <- paste0(path_base, ".svg")
+    ok <- tryCatch(
+      .save_widget_svg(html, out, width, height, delay),
+      error = function(e) {
+        if (verbose) message("    svg failed: ", conditionMessage(e))
+        FALSE
+      }
+    )
+    if (isTRUE(ok)) written <- c(written, out)
+  }
+
+  written
+}
+
+
+#' Re-encode a PNG into another raster/PDF format via magick
+#'
+#' PDF is written as a rasterised image wrapped in a PDF page (not vector).
+#'
+#' @param src PNG source path.
+#' @param out Destination path.
+#' @param ext Target extension ("jpg", "jpeg", "pdf").
+#' @param verbose Whether to message on failure.
+#' @return TRUE on success, FALSE otherwise.
+#' @keywords internal
+#' @noRd
+.convert_raster <- function(src, out, ext, verbose = TRUE) {
+  fmt <- if (ext %in% c("jpg", "jpeg")) "jpeg" else ext
+  tryCatch(
+    {
+      magick::image_write(magick::image_read(src), out, format = fmt)
+      file.exists(out) && file.info(out)$size > 0
+    },
+    error = function(e) {
+      if (verbose) message("    ", ext, " failed: ", conditionMessage(e))
+      FALSE
+    }
+  )
+}
+
+
+#' Extract an htmlwidget's rendered SVG element into a standalone .svg file
+#'
+#' Plotly widgets go through `Plotly.toImage` so title/axis/legend overlays
+#' end up inlined as SVG text; other widgets use the raw SVG outerHTML.
+#'
+#' @param html Path to the saved widget HTML.
+#' @param out Path to write the SVG to.
+#' @param width,height Viewport size the widget renders into.
+#' @param delay Seconds to wait after navigation before extracting.
+#' @return TRUE on success, FALSE if the widget has no SVG or writes fail.
+#' @keywords internal
+#' @noRd
+.save_widget_svg <- function(html, out, width, height, delay) {
+  b <- chromote::ChromoteSession$new()
+  on.exit(b$close(), add = TRUE)
+  b$Emulation$setDeviceMetricsOverride(
+    width = as.integer(width), height = as.integer(height),
+    deviceScaleFactor = 1, mobile = FALSE
+  )
+  b$Page$navigate(paste0("file://", html))
+  Sys.sleep(delay)
+  js <- sprintf(
+    "(async function() {
+       var gd = document.querySelector('.js-plotly-plot');
+       if (gd && window.Plotly) {
+         var url = await Plotly.toImage(gd, {format: 'svg', width: %d, height: %d});
+         return decodeURIComponent(url.replace(/^data:image\\/svg\\+xml,/, ''));
+       }
+       var s = document.querySelector('svg');
+       if (!s) return null;
+       // Drop tooltip <title> nodes: networkD3 puts <foreignObject><body>
+       // inside them which strict SVG rasterisers reject, and they render
+       // nothing standalone anyway.
+       s.querySelectorAll('title').forEach(function(t) { t.remove(); });
+       if (!s.getAttribute('xmlns')) s.setAttribute('xmlns', 'http://www.w3.org/2000/svg');
+       return s.outerHTML;
+     })()",
+    as.integer(width), as.integer(height)
+  )
+  res <- b$Runtime$evaluate(js, returnByValue = TRUE, awaitPromise = TRUE)
+  svg_str <- res$result$value
+  if (is.null(svg_str) || !nzchar(svg_str)) return(FALSE)
+  writeLines(svg_str, out)
+  file.exists(out) && file.info(out)$size > 0
+}
+
+
+#' Baseline (non-stratified) drug labels available for a species
+#' @keywords internal
+#' @noRd
+.export_ml_drugs <- function(perf_data, code) {
+  if (is.null(perf_data) || !nrow(perf_data)) {
+    return(character(0))
+  }
+  perf_data |>
+    dplyr::filter(
+      normalize_species(.data$species) %in% normalize_species(code)
+    ) |>
+    dplyr::filter(is.na(.data$strat_label) | !nzchar(.data$strat_label)) |>
+    dplyr::filter(.data$drug_label == "drug") |>
+    dplyr::pull(.data$drug_or_class) |>
+    unique() |>
+    sort()
+}
+
+
+#' Baseline top-feature drug labels for a species (feature-importance defaults)
+#' @keywords internal
+#' @noRd
+.export_tf_drugs <- function(top_features, code,
+                             scale = "genes",
+                             subtype = c("counts", "binary")) {
+  if (is.null(top_features) || !nrow(top_features)) {
+    return(character(0))
+  }
+  top_features |>
+    dplyr::filter(
+      normalize_species(.data$species) %in% normalize_species(code)
+    ) |>
+    dplyr::filter(.data$feature_type %in% scale) |>
+    dplyr::filter(.data$feature_subtype %in% subtype) |>
+    dplyr::filter(is.na(.data$strat_label) | !nzchar(.data$strat_label)) |>
+    dplyr::filter(.data$drug_label == "drug") |>
+    dplyr::pull(.data$drug_or_class) |>
+    unique() |>
+    sort()
+}
+
+
+#' Crop uniform-white margins from a raster file in place
+#'
+#' No-op when `magick` isn't installed or the trim errors.
+#'
+#' @param path Raster file to overwrite in place.
+#' @param verbose Whether to message on failure.
+#' @keywords internal
+#' @noRd
+.trim_raster <- function(path, verbose = TRUE) {
+  tryCatch(
+    magick::image_write(magick::image_trim(magick::image_read(path)), path),
+    error = function(e) {
+      if (verbose) message("    trim failed: ", conditionMessage(e))
+    }
+  )
+  invisible()
+}
+
+
+#' Fit a networkD3 forceNetwork to its container after simulation settles
+#'
+#' Waits for the force simulation, then makes the SVG fill its container and
+#' rewrites the viewBox to the graph's bbox. Export-only.
+#'
+#' @param widget A `networkD3::forceNetwork` htmlwidget.
+#' @return The widget with the onRender hook attached.
+#' @keywords internal
+#' @noRd
+.fit_network_to_content <- function(widget) {
+  htmlwidgets::onRender(widget, "
+    function(el, x) {
+      setTimeout(function() {
+        try {
+          var svg = el.querySelector('svg');
+          svg.setAttribute('width', '100%');
+          svg.setAttribute('height', '100%');
+          var bb = svg.getBBox();
+          var p = 40;
+          svg.setAttribute('viewBox',
+            (bb.x - p) + ' ' + (bb.y - p) + ' ' +
+            (bb.width + 2*p) + ' ' + (bb.height + 2*p));
+        } catch (e) {}
+      }, 1300);
+    }
+  ")
+}
+
+
+#' Assemble the list of figures to export
+#'
+#' Returns a list of specs; each is `list(group, name, build)` where `build()`
+#' lazily produces the widget. Grouping mirrors the dashboard tabs. Global
+#' overviews live under `_overview` / `_across_species`; everything else is
+#' filed under its species folder. Selections replicate the dashboard defaults.
+#'
+#' @keywords internal
+#' @noRd
+.exportPlanSpecs <- function(perf_data, top_features,
+                             ml_species, meta_species,
+                             results_root, amrdata_root,
+                             top_n_features = 10, network_top_n = 5) {
+  specs <- list()
+  add <- function(group, name, build,
+                  width = NULL, height = NULL, trim = FALSE) {
+    specs[[length(specs) + 1]] <<- list(
+      group = group, name = name, build = build,
+      width = width, height = height, trim = trim
+    )
+  }
+
+  model_scale_all <- c("genes", "proteins", "domains", "cogs", "args")
+  fi_scale <- "genes"
+  fi_subtype <- c("counts", "binary")
+
+  # Defensive: drop species with no code. Blank/NA codes share a species_label
+  # folder with a real species, so their per-species specs would collide with
+  # (and overwrite) the real ones. Guaranteeing unique output paths here keeps
+  # the plan robust regardless of how the caller derived `ml_species`.
+  valid <- !is.na(ml_species$code) & nzchar(ml_species$code)
+  ml_species <- list(
+    code = ml_species$code[valid], label = ml_species$label[valid]
+  )
+
+  # ---- Global performance overviews (all species at once) ----
+  if (nrow(perf_data)) {
+    add("_overview", "mcc_strip_by_species_scale", function() {
+      makeMCCStripPlot(perf_data, selected_drug_class = "all")
+    })
+    add("_overview", "mcc_heatmap_overview", function() {
+      makeMCCHeatmap(perf_data, selected_drug_class = "all")
+    })
+  }
+
+  # ---- Cross-species feature importance (across_bug view) ----
+  if (nrow(top_features) && length(ml_species$code)) {
+    pooled_drugs <- .export_tf_drugs(
+      top_features, ml_species$code, fi_scale, fi_subtype
+    )
+    if (length(pooled_drugs)) {
+      drug <- if ("GEN" %in% pooled_drugs) "GEN" else pooled_drugs[1]
+      add("_across_species", paste0("feature_importance_", drug), function() {
+        makeFeatureImportancePlot(
+          top_features, ml_species$code, drug,
+          fi_scale, fi_subtype, top_n_features, "across_bug",
+          amrdata_root = amrdata_root, results_root = results_root
+        )
+      })
+    }
+  }
+
+  # ---- Per ML species: performance, feature importance, holdouts, network ----
+  for (i in seq_along(ml_species$code)) {
+    code <- ml_species$code[i]
+    folder <- ml_species$label[i]
+
+    # Model performance box/point plot.
+    drugs <- .export_ml_drugs(perf_data, code)
+    drug_sel <- if ("GEN" %in% drugs) {
+      "GEN"
+    } else if (length(drugs)) {
+      drugs[1]
+    } else {
+      NULL
+    }
+    local({
+      code <- code
+      drug_sel <- drug_sel
+      add(folder, "model_performance", function() {
+        makeModelPerformancePlot(
+          perf_data, code, model_scale_all, "binary", "mcc",
+          "all", drug_sel
+        )
+      })
+    })
+
+    # Feature importance across drugs for this species.
+    tf_drugs <- .export_tf_drugs(top_features, code, fi_scale, fi_subtype)
+    fi_drugs <- intersect(c("OXA", "PEN", "MET"), tf_drugs)
+    if (!length(fi_drugs)) {
+      fi_drugs <- utils::head(tf_drugs, min(3, length(tf_drugs)))
+    }
+    if (length(fi_drugs)) {
+      local({
+        code <- code
+        fi_drugs <- fi_drugs
+        add(folder, "feature_importance_across_drugs", function() {
+          makeFeatureImportancePlot(
+            top_features, code, fi_drugs,
+            fi_scale, fi_subtype, top_n_features, "across_drug",
+            amrdata_root = amrdata_root, results_root = results_root
+          )
+        })
+        # COG category barplot from annotation-enriched features.
+        add(folder, "cog_categories", function() {
+          tf <- top_features |>
+            dplyr::filter(
+              is.na(.data$strat_label) | !nzchar(.data$strat_label)
+            ) |>
+            dplyr::filter(!isTRUE(.data$cross_test)) |>
+            dplyr::filter(
+              normalize_species(.data$species) %in% normalize_species(code)
+            ) |>
+            dplyr::filter(.data$drug_or_class %in% fi_drugs)
+          if (!nrow(tf)) {
+            return(makeCogBarChart(NULL))
+          }
+          enriched <- dplyr::bind_rows(lapply(unique(tf$species), function(sp) {
+            enrich_with_annotations(
+              tf[tf$species == sp, ],
+              species_code = sp, results_root = results_root
+            )
+          }))
+          makeCogBarChart(enriched)
+        })
+      })
+    }
+
+    # Cross-model holdouts (country + time strata).
+    holdout_drugs <- getHoldoutsDrugChoices(perf_data, code)
+    holdout_drug <- if (length(holdout_drugs)) holdout_drugs[1] else NULL
+    local({
+      code <- code
+      holdout_drug <- holdout_drug
+      add(folder, "holdout_ridge_country", function() {
+        makeCrossModelRidgePlot(perf_data, code, "country")
+      })
+      add(folder, "holdout_ridge_time", function() {
+        makeCrossModelRidgePlot(perf_data, code, "time")
+      })
+      if (!is.null(holdout_drug)) {
+        add(
+          folder, paste0("holdout_performance_country_", holdout_drug),
+          function() {
+            makeCrossModelPerformancePlot(
+              perf_data, code, holdout_drug, "country"
+            )
+          }
+        )
+        add(
+          folder, paste0("holdout_performance_time_", holdout_drug),
+          function() {
+            makeCrossModelPerformancePlot(
+              perf_data, code, holdout_drug, "time"
+            )
+          }
+        )
+        add(
+          folder, paste0("holdout_feature_importance_country_", holdout_drug),
+          function() {
+            makeCrossModelFeatureImportancePlot(
+              top_features, code, holdout_drug, "country", top_n_features
+            )
+          }
+        )
+      }
+    })
+
+    # Drug-feature network. Square viewport since the graph settles roughly
+    # square; .fit_network_to_content() then fits the SVG to the actual bbox.
+    local({
+      code <- code
+      add(folder, "drug_feature_network", function() {
+        makeDrugFeatureNetwork(
+          top_features, code,
+          top_n = network_top_n,
+          include_clusters = FALSE, include_cogs = FALSE,
+          results_root = results_root
+        ) |> .fit_network_to_content()
+      }, width = 1600, height = 1600, trim = TRUE)
+    })
+  }
+
+  # ---- Per metadata species: distributions + sankey ----
+  for (sp in meta_species) {
+    folder <- sp
+    local({
+      sp <- sp
+      folder <- folder
+      meta_raw <- function() .read_metadata_for_bug(sp, results_root)
+
+      add(folder, "metadata_data_availability", function() {
+        makeDatAvailabilityPlot(meta_raw())
+      })
+      add(folder, "metadata_geographic", function() {
+        data <- meta_raw() |>
+          dplyr::filter(.data$genome.isolation_country != "") |>
+          .add_evidence_column() |>
+          dplyr::group_by(
+            .data$genome.isolation_country, .data$genome_drug.antibiotic
+          ) |>
+          dplyr::summarize(count = dplyr::n(), .groups = "drop") |>
+          dplyr::group_by(.data$genome.isolation_country) |>
+          dplyr::summarise(count = sum(.data$count), .groups = "drop")
+        makeGeoChloroPlot(data)
+      })
+      add(folder, "metadata_resistance_over_time", function() {
+        data <- meta_raw() |>
+          .add_evidence_column() |>
+          dplyr::filter(!is.na(.data$genome.collection_year)) |>
+          dplyr::group_by(
+            .data$genome_drug.antibiotic,
+            .data$genome_drug.resistant_phenotype,
+            .data$genome.isolation_country,
+            .data$genome.collection_year
+          ) |>
+          dplyr::summarize(n = dplyr::n(), .groups = "drop")
+        makeTimeSeriesAMRPlot(data, "all")
+      })
+      add(folder, "metadata_hosts", function() {
+        data <- meta_raw() |>
+          dplyr::filter(.data$genome.host_common_name != "") |>
+          .add_evidence_column()
+        makeHostIsolatePlot(data)
+      })
+      add(folder, "metadata_isolation_sources", function() {
+        data <- meta_raw() |>
+          dplyr::filter(.data$genome.host_common_name != "") |>
+          .add_evidence_column()
+        makeIsolationSourcesPlot(data)
+      })
+      add(folder, "metadata_resistance_sankey", function() {
+        meta <- meta_raw()
+        classes <- if (nrow(meta) && "drug_class" %in% names(meta)) {
+          meta |>
+            dplyr::filter(!is.na(.data$drug_class)) |>
+            dplyr::count(.data$drug_class, name = "n") |>
+            dplyr::arrange(dplyr::desc(.data$n)) |>
+            dplyr::slice_head(n = 3) |>
+            dplyr::pull(.data$drug_class)
+        } else {
+          NULL
+        }
+        makeMetadataSankey(meta, drug_classes = classes)
+      })
+    })
+  }
+
+  specs
+}
+
+
+#' Export all amRviz dashboard visualizations to static image files
+#'
+#' Renders every amRviz figure (metadata, model performance, feature
+#' importance, cross-model holdouts, drug-feature networks) to files, without
+#' launching the Shiny app. One figure set per species, with global overviews
+#' under `_overview` / `_across_species`. Selections mirror the dashboard's
+#' defaults; `top_n_features` and `network_top_n` override the corresponding
+#' sliders.
+#'
+#' Widgets are driven by a headless Chrome: `png` is a webshot2 snapshot;
+#' `jpg` and `pdf` are re-encodes of that PNG via \pkg{magick} (identical
+#' crop and dimensions); `svg` is extracted from the DOM via \pkg{chromote},
+#' using `Plotly.toImage` for plotly so titles and labels survive.
+#'
+#' @param output_dir Directory to write figures into; created if needed. Files
+#'   are organised as `output_dir/<species>/<panel>.<ext>`, with global
+#'   overviews under `output_dir/_overview` and `output_dir/_across_species`.
+#' @param formats Any of `"png"`, `"pdf"`, `"jpg"`, `"svg"`.
+#' @param results_root Directory of amRml model outputs (per-species subdirs
+#'   of `*_perf.parquet` / `*_top_features.parquet` / `metadata.parquet`).
+#'   `NULL` uses the packaged demo data.
+#' @param amrdata_root Directory of amRdata annotation parquets (for COG
+#'   enrichment). `NULL` tries `~/amRdata/data`, else falls back to unenriched.
+#' @param species Optional character vector restricting which species folders
+#'   to export.
+#' @param top_n_features,network_top_n Feature counts for the
+#'   feature-importance panels and the drug-feature network.
+#' @param width,height Viewport size in pixels.
+#' @param scale Device-pixel multiplier for the PNG render; higher = sharper.
+#' @param delay Seconds to wait for each widget's JavaScript to settle.
+#' @param verbose Per-figure progress messages.
+#'
+#' @return Invisibly, a data frame with one row per figure: `group`, `name`,
+#'   `written` (file count), `ok`.
+#' @export
+#' @examples
+#' if (interactive()) {
+#'   # Export the packaged demo figures as PNG + PDF into ./amRviz_exports
+#'   exportAMRVisualizations()
+#'
+#'   # Your own results, PNG only, one species, more features per panel
+#'   exportAMRVisualizations(
+#'     output_dir = "figs",
+#'     formats = "png",
+#'     results_root = "~/my_amRml_results",
+#'     species = "Shigella_flexneri",
+#'     top_n_features = 25,
+#'     network_top_n = 10
+#'   )
+#' }
+exportAMRVisualizations <- function(output_dir = "amRviz_exports",
+                                    formats = c("png", "pdf"),
+                                    results_root = NULL,
+                                    amrdata_root = NULL,
+                                    species = NULL,
+                                    top_n_features = 10,
+                                    network_top_n = 5,
+                                    width = 1200,
+                                    height = 800,
+                                    scale = 2,
+                                    delay = 1.5,
+                                    verbose = TRUE) {
+  # Validate dependencies up front with actionable messages.
+  for (pkg in c("htmlwidgets", "webshot2", "chromote", "magick")) {
+    if (!requireNamespace(pkg, quietly = TRUE)) {
+      stop(
+        "Package '", pkg, "' is required for exportAMRVisualizations(). ",
+        "Install it with install.packages('", pkg, "').",
+        call. = FALSE
+      )
+    }
+  }
+  chrome_ok <- tryCatch(
+    nzchar(chromote::find_chrome()),
+    error = function(e) FALSE
+  )
+  if (!isTRUE(chrome_ok)) {
+    stop(
+      "A Chrome/Chromium browser is required to render the figures but none ",
+      "was found. Install Google Chrome or Chromium (webshot2/chromote use it ",
+      "to photograph the interactive plots).",
+      call. = FALSE
+    )
+  }
+
+  formats <- tolower(formats)
+  formats[formats == "jpeg"] <- "jpg"
+  valid <- c("png", "pdf", "jpg", "svg")
+  bad <- setdiff(formats, valid)
+  if (length(bad)) {
+    stop(
+      "Unsupported format(s): ", paste(bad, collapse = ", "),
+      ". Choose from: ", paste(valid, collapse = ", "), ".",
+      call. = FALSE
+    )
+  }
+
+  if (!is.numeric(scale) || length(scale) != 1 || is.na(scale) || scale <= 0) {
+    stop("`scale` must be a single positive number.", call. = FALSE)
+  }
+
+  for (nm in c("top_n_features", "network_top_n")) {
+    v <- get(nm)
+    if (!is.numeric(v) || length(v) != 1 || is.na(v) || v < 1) {
+      stop("`", nm, "` must be a single positive number.", call. = FALSE)
+    }
+  }
+
+  # Default amrdata_root: ~/amRdata/data when present (mirrors the dashboard).
+  if (is.null(amrdata_root)) {
+    default_amrdata <- file.path(path.expand("~"), "amRdata", "data")
+    if (dir.exists(default_amrdata)) amrdata_root <- default_amrdata
+  }
+
+  # Load model results (user results_root, else packaged demo data).
+  species_dirs <- if (!is.null(results_root) && nzchar(results_root)) {
+    unname(listAmRmlSpeciesFolders(results_root))
+  } else {
+    NULL
+  }
+  perf_data <- loadMLResults(results_root, species_dirs, verbose = verbose)
+  top_features <- loadTopFeat(results_root, species_dirs, verbose = verbose)
+
+  # ML species: (code, folder label) pairs, excluding cross/MDR pseudo-species.
+  ml_species <- list(code = character(0), label = character(0))
+  if (nrow(perf_data) &&
+    all(c("species", "species_label") %in% names(perf_data))) {
+    # Drop rows with no species code: some parquets carry NA/blank species,
+    # which would otherwise yield a phantom species whose folder collides with
+    # (and overwrites) a real one. The dashboard drops these implicitly via
+    # sort(); we do it explicitly.
+    pairs <- perf_data |>
+      dplyr::filter(!is.na(.data$species) & nzchar(.data$species)) |>
+      dplyr::filter(!(.data$species %in% c("cross", "MDR"))) |>
+      dplyr::distinct(.data$species, .data$species_label) |>
+      dplyr::arrange(.data$species_label, .data$species)
+    ml_species <- list(
+      code = as.character(pairs$species),
+      label = as.character(pairs$species_label)
+    )
+  }
+
+  # Metadata species: subdirectories that hold a metadata.parquet.
+  scan_root <- if (!is.null(results_root) && nzchar(results_root)) {
+    results_root
+  } else {
+    system.file("extdata", package = "amRviz")
+  }
+  meta_species <- character(0)
+  if (nzchar(scan_root)) {
+    for (d in list.dirs(scan_root, full.names = TRUE, recursive = FALSE)) {
+      if (file.exists(file.path(d, "metadata.parquet"))) {
+        meta_species <- c(meta_species, basename(d))
+      }
+    }
+    meta_species <- sort(unique(meta_species))
+  }
+
+  # Optional species filter (matched against folder labels).
+  if (!is.null(species)) {
+    keep <- ml_species$label %in% species | ml_species$code %in% species
+    ml_species <- list(
+      code = ml_species$code[keep], label = ml_species$label[keep]
+    )
+    meta_species <- meta_species[meta_species %in% species]
+  }
+
+  if (!length(ml_species$code) && !length(meta_species)) {
+    stop(
+      "No species found to export. Check `results_root` / `species`, or omit ",
+      "them to use the packaged demo data.",
+      call. = FALSE
+    )
+  }
+
+  specs <- .exportPlanSpecs(
+    perf_data, top_features, ml_species, meta_species,
+    results_root, amrdata_root,
+    top_n_features = top_n_features, network_top_n = network_top_n
+  )
+
+  if (!dir.exists(output_dir)) {
+    dir.create(output_dir, recursive = TRUE)
+  }
+
+  if (verbose) {
+    message(
+      "Exporting ", length(specs), " figures (",
+      paste(formats, collapse = ", "), ") to ",
+      normalizePath(output_dir, mustWork = FALSE)
+    )
+  }
+
+  results <- vector("list", length(specs))
+  for (i in seq_along(specs)) {
+    spec <- specs[[i]]
+    group_dir <- file.path(output_dir, spec$group)
+    if (!dir.exists(group_dir)) dir.create(group_dir, recursive = TRUE)
+    path_base <- file.path(group_dir, spec$name)
+
+    if (verbose) {
+      message(
+        "  [", i, "/", length(specs), "] ",
+        spec$group, "/", spec$name
+      )
+    }
+
+    written <- tryCatch(
+      {
+        widget <- spec$build()
+        .exportWidgetFile(
+          widget, path_base, formats,
+          width = spec$width %||% width,
+          height = spec$height %||% height,
+          scale = scale, trim = isTRUE(spec$trim),
+          delay = delay, verbose = verbose
+        )
+      },
+      error = function(e) {
+        if (verbose) message("    build failed: ", conditionMessage(e))
+        character(0)
+      }
+    )
+
+    results[[i]] <- data.frame(
+      group = spec$group, name = spec$name,
+      written = length(written), ok = length(written) > 0,
+      stringsAsFactors = FALSE
+    )
+  }
+
+  summary_df <- do.call(rbind, results)
+  if (verbose) {
+    n_ok <- sum(summary_df$ok)
+    n_files <- sum(summary_df$written)
+    message(
+      "Done: ", n_ok, "/", nrow(summary_df),
+      " figures rendered, ", n_files, " files written."
+    )
+    failed <- summary_df[!summary_df$ok, , drop = FALSE]
+    if (nrow(failed)) {
+      message(
+        "  No output for: ",
+        paste(
+          paste0(failed$group, "/", failed$name),
+          collapse = ", "
+        )
+      )
+    }
+  }
+
+  invisible(summary_df)
+}
diff --git a/R/plots_crossmodel.R b/R/plots_crossmodel.R
index 5a45adf..a25dbf1 100644
--- a/R/plots_crossmodel.R
+++ b/R/plots_crossmodel.R
@@ -50,9 +50,9 @@ makeCrossModelFeatureImportancePlot <- function(
   }
 
   vi_wider <- features_df |>
-    dplyr::select("Variable", "Importance", !!rlang::sym(strat_col)) |>
+    dplyr::select("Variable", "Importance", dplyr::all_of(strat_col)) |>
     tidyr::pivot_wider(
-      names_from = strat_col,
+      names_from = dplyr::all_of(strat_col),
       values_from = "Importance",
       values_fn = mean
     )
@@ -109,8 +109,7 @@ makeCrossModelFeatureImportancePlot <- function(
 #' @noRd
 makeCrossModelRidgePlot <- function(perf_data, bug, cross_model) {
   if (is.null(perf_data) || !is.data.frame(perf_data) || !nrow(perf_data)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No data available", x = 0)))
+    return(plotly_placeholder("No data available"))
   }
 
   strat <- if (cross_model == "country") "country" else "year"
@@ -123,8 +122,7 @@ makeCrossModelRidgePlot <- function(perf_data, bug, cross_model) {
     dplyr::filter(.data$drug_label == "drug_class")
 
   if (!nrow(df)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No data for selection", x = 0)))
+    return(plotly_placeholder("No data for selection"))
   }
 
   # Label: Same = trained & tested on same stratum, Different = cross-tested
@@ -198,13 +196,8 @@ makeCrossModelRidgePlot <- function(perf_data, bug, cross_model) {
 #' @keywords internal
 #' @noRd
 makeCrossModelPerformancePlot <- function(perf_data, bug, drug, cross_model) {
-  placeholder <- function(msg) {
-    plotly::plot_ly() |>
-      plotly::layout(title = list(text = msg, x = 0))
-  }
-
   if (is.null(perf_data) || !is.data.frame(perf_data) || !nrow(perf_data)) {
-    return(placeholder("No data available"))
+    return(plotly_placeholder("No data available"))
   }
 
   strat <- if (cross_model == "country") "country" else "year"
@@ -218,14 +211,14 @@ makeCrossModelPerformancePlot <- function(perf_data, bug, drug, cross_model) {
     dplyr::filter(.data$strat_label == strat)
 
   if (!nrow(df)) {
-    return(placeholder("No data for selection"))
+    return(plotly_placeholder("No data for selection"))
   }
 
   # A cross-stratum grid needs at least two trained-on strata to compare.
   n_strata <- dplyr::n_distinct(df$strat_value)
   if (n_strata < 2) {
     unit <- if (cross_model == "country") "countries" else "time periods"
-    return(placeholder(paste0(
+    return(plotly_placeholder(paste0(
       "Not enough data to compare: needs models trained on 2+ ", unit,
       " (this selection has ", n_strata, ")."
     )))
@@ -254,7 +247,7 @@ makeCrossModelPerformancePlot <- function(perf_data, bug, drug, cross_model) {
     as.matrix()
 
   if (!length(models_performance)) {
-    return(placeholder("No data for selection"))
+    return(plotly_placeholder("No data for selection"))
   }
 
   plotly::plot_ly(
diff --git a/R/plots_featureimportance.R b/R/plots_featureimportance.R
index 9d3bb24..b7c9b4c 100644
--- a/R/plots_featureimportance.R
+++ b/R/plots_featureimportance.R
@@ -299,8 +299,7 @@ makeFeatureImportancePlot <- function(
 makeCogBarChart <- function(enriched_tbl, top_n = 15) {
   if (is.null(enriched_tbl) || !nrow(enriched_tbl) ||
     !"COG" %in% names(enriched_tbl)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No annotations available", x = 0)))
+    return(plotly_placeholder("No annotations available"))
   }
 
   # Split the comma-separated COG cells and count occurrences per Variable.
@@ -310,8 +309,7 @@ makeCogBarChart <- function(enriched_tbl, top_n = 15) {
     dplyr::distinct()
 
   if (!nrow(cog_df)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No COGs in selection", x = 0)))
+    return(plotly_placeholder("No COGs in selection"))
   }
 
   rows <- do.call(rbind, lapply(seq_len(nrow(cog_df)), function(i) {
diff --git a/R/plots_modelperf.R b/R/plots_modelperf.R
index 4f15a56..0a061ce 100644
--- a/R/plots_modelperf.R
+++ b/R/plots_modelperf.R
@@ -24,17 +24,17 @@ makeModelPerformancePlot <- function(
   amr_drug_class, amr_drug
 ) {
   if (is.null(data) || !is.data.frame(data) || !nrow(data)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No data available", x = 0)))
+    return(plotly_placeholder("No data available"))
   }
 
-  # Filter to baseline models (strat_label is NA = no country/year
-  # stratification); drop cross-tested rows, which ship in their own files.
+  # Baseline models only (strat_label NA); drop cross-tested rows (own files).
+  # Drop NA metric rows so plotly's box trace doesn't warn about dropped obs.
   df <- data |>
     dplyr::filter(normalize_species(.data$species) %in% normalize_species(bug)) |>
     dplyr::filter(.data$feature_type %in% model_scale) |>
     dplyr::filter(.data$feature_subtype %in% data_type) |>
-    dplyr::filter(is.na(.data$strat_label) | !nzchar(.data$strat_label))
+    dplyr::filter(is.na(.data$strat_label) | !nzchar(.data$strat_label)) |>
+    dplyr::filter(!is.na(.data[[metrics]]))
   if ("cross_test" %in% names(df)) {
     df <- dplyr::filter(df, !.data$cross_test)
   }
@@ -50,8 +50,7 @@ makeModelPerformancePlot <- function(
   }
 
   if (!nrow(df)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = list(text = "No data for current selection", x = 0)))
+    return(plotly_placeholder("No data for current selection"))
   }
 
   # Normalize species and set ESKAPE factor order
@@ -209,7 +208,7 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
                              selected_drug = NULL) {
   df <- .prep_mcc_data(data)
   if (is.null(df)) {
-    return(plotly::plot_ly() |> plotly::layout(title = "No data available"))
+    return(plotly_placeholder("No data available"))
   }
 
   sl <- .scale_label_map()
@@ -227,8 +226,7 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
     )
 
   if (!nrow(df)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = "No data for selected filters"))
+    return(plotly_placeholder("No data for selected filters"))
   }
 
   # Highlight selected drug or drug class. Priority: specific drug > class.
@@ -262,6 +260,9 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
       )
     )
 
+  # Keep alpha/size on geom_jitter only. In the top-level aes() they'd also
+  # apply to geom_boxplot's line width, triggering ggplot2 3.4's size-for-lines
+  # deprecation.
   g <- ggplot2::ggplot(
     df,
     ggplot2::aes(
@@ -269,8 +270,6 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
       y = .data$mcc,
       color = .data$scale_label,
       fill = .data$scale_label,
-      alpha = .data$pt_alpha,
-      size = .data$pt_size,
       text = paste0(
         "Drug/class: ", .data$drug_or_class,
         "\nMCC: ", round(.data$mcc, 3),
@@ -282,7 +281,10 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
       alpha = 0.3, outlier.shape = NA,
       width = 0.5, linewidth = 0.4
     ) +
-    ggplot2::geom_jitter(width = 0.15) +
+    ggplot2::geom_jitter(
+      ggplot2::aes(alpha = .data$pt_alpha, size = .data$pt_size),
+      width = 0.15
+    ) +
     ggplot2::geom_hline(
       yintercept = 0, linetype = "dashed",
       color = "gray50", linewidth = 0.4
@@ -328,12 +330,11 @@ makeMCCStripPlot <- function(data, selected_drug_class = NULL,
 makeMCCHeatmap <- function(data, selected_drug_class = NULL) {
   df <- .prep_mcc_data(data)
   if (is.null(df)) {
-    return(plotly::plot_ly() |> plotly::layout(title = "No data available"))
+    return(plotly_placeholder("No data available"))
   }
   df <- df |> dplyr::filter(.data$drug_label == "drug_class")
   if (!nrow(df)) {
-    return(plotly::plot_ly() |>
-      plotly::layout(title = "No drug-class data available"))
+    return(plotly_placeholder("No drug-class data available"))
   }
 
   # Drug-class order: most-represented first (bottom of y = first row)
diff --git a/R/utils_misc.R b/R/utils_misc.R
index 84ec5bc..3cb3682 100644
--- a/R/utils_misc.R
+++ b/R/utils_misc.R
@@ -1,4 +1,23 @@
-# Misc cross-cutting helpers: species-code handling, choice derivation.
+# Misc cross-cutting helpers.
+
+
+#' Empty plotly placeholder with an explanatory title
+#'
+#' Returned by plots_* functions when there's nothing to draw. Sets an explicit
+#' trace type so plotly doesn't warn about guessing one at render.
+#'
+#' @param msg Message shown as the plot title.
+#' @return A `plotly` htmlwidget.
+#' @keywords internal
+#' @noRd
+plotly_placeholder <- function(msg) {
+  plotly::plot_ly(type = "scatter", mode = "markers") |>
+    plotly::layout(
+      title = list(text = msg, x = 0),
+      xaxis = list(visible = FALSE),
+      yaxis = list(visible = FALSE)
+    )
+}
 
 
 # Species code regex (note: Esp. includes escaped period)
diff --git a/README.Rmd b/README.Rmd
index 605a8b5..ba06c9d 100644
--- a/README.Rmd
+++ b/README.Rmd
@@ -35,6 +35,7 @@ This is the final package in the **AMR package suite**, [JRaviLab/amR](https://g
 - **Cross-model analysis**: Compare models trained on different stratifications (country, year)
 - **Dynamic species selection**: Dropdowns automatically populate from loaded data — no hardcoded species lists
 - **Demo mode**: Ships with example *Shigella flexneri* data; swap in your own amRml output with one argument
+- **Headless figure export**: Render every visualization to PNG/PDF/JPG files without launching the dashboard, with one call to `exportAMRVisualizations()`
 
 ## Installation
 
@@ -68,6 +69,36 @@ launchAMRDashboard(results_root = "/path/to/your/amRml/results")
 
 The dashboard will open in your default web browser. Species dropdowns will populate automatically from whichever data is loaded.
 
+## Export figures without the dashboard
+
+If you'd rather not run the interactive dashboard, `exportAMRVisualizations()` renders every visualization to static image files in a single call — handy for reports, batch pipelines, or a quick look at all the figures at once.
+
+```{r, eval = FALSE}
+library(amRviz)
+
+# Export the bundled demo figures as PNG + PDF into ./amRviz_exports/
+exportAMRVisualizations()
+
+# Your own results, PNG + JPG, a single species
+exportAMRVisualizations(
+  output_dir   = "figures",
+  formats      = c("png", "jpg"),
+  results_root = "/path/to/your/amRml/results",
+  species      = "Shigella_flexneri"
+)
+```
+
+One figure set is produced per species using the same default selections the dashboard opens with, organized as `output_dir/<species>/<panel>.<ext>`. Cross-species overviews (the performance heatmaps and the across-species feature-importance panel) are written once under `_overview/` and `_across_species/`.
+
+You can adjust how many features appear with `top_n_features` (feature-importance panels) and `network_top_n` (the drug-feature network), and control raster resolution with `scale`:
+
+```{r, eval = FALSE}
+exportAMRVisualizations(top_n_features = 25, network_top_n = 10, scale = 3)
+```
+
+- **Formats**: all four (`png`, `jpg`, `pdf`, `svg`) are supported. `png` is a headless-Chrome screenshot; `jpg` and `pdf` are re-encodes of that PNG, so every raster format shares the same crop and dimensions. `svg` is extracted from the rendered DOM (with all text inlined), giving a scalable, vector figure — the best choice for publication.
+- **Requirements**: every plot is an interactive htmlwidget, so export drives each one with a headless Chrome via the `webshot2` and `chromote` packages, then reformats with `magick`. Install Google Chrome or Chromium if you don't already have one; the function stops early with a clear message if no browser is found.
+
 ## Usage
 
 ### Dashboard navigation
diff --git a/README.md b/README.md
index 7094273..1beee81 100644
--- a/README.md
+++ b/README.md
@@ -34,6 +34,9 @@ This is the final package in the **AMR package suite**,
     loaded data — no hardcoded species lists
 -   **Demo mode**: Ships with example *Shigella flexneri* data; swap in
     your own amRml output with one argument
+-   **Headless figure export**: Render every visualization to PNG/PDF/JPG
+    files without launching the dashboard, with one call to
+    `exportAMRVisualizations()`
 
 ## Installation
 
@@ -67,6 +70,54 @@ launchAMRDashboard(results_root = "/path/to/your/amRml/results")
 The dashboard will open in your default web browser. Species dropdowns
 will populate automatically from whichever data is loaded.
 
+## Export figures without the dashboard
+
+If you’d rather not run the interactive dashboard,
+`exportAMRVisualizations()` renders every visualization to static image
+files in a single call — handy for reports, batch pipelines, or a quick
+look at all the figures at once.
+
+``` r
+library(amRviz)
+
+# Export the bundled demo figures as PNG + PDF into ./amRviz_exports/
+exportAMRVisualizations()
+
+# Your own results, PNG + JPG, a single species
+exportAMRVisualizations(
+  output_dir   = "figures",
+  formats      = c("png", "jpg"),
+  results_root = "/path/to/your/amRml/results",
+  species      = "Shigella_flexneri"
+)
+```
+
+One figure set is produced per species using the same default selections
+the dashboard opens with, organized as
+`output_dir/<species>/<panel>.<ext>`. Cross-species overviews (the
+performance heatmaps and the across-species feature-importance panel)
+are written once under `_overview/` and `_across_species/`.
+
+You can adjust how many features appear with `top_n_features`
+(feature-importance panels) and `network_top_n` (the drug-feature
+network), and control raster resolution with `scale`:
+
+``` r
+exportAMRVisualizations(top_n_features = 25, network_top_n = 10, scale = 3)
+```
+
+-   **Formats**: all four (`png`, `jpg`, `pdf`, `svg`) are supported.
+    `png` is a headless-Chrome screenshot; `jpg` and `pdf` are re-encodes
+    of that PNG, so every raster format shares the same crop and
+    dimensions. `svg` is extracted from the rendered DOM (with all text
+    inlined), giving a scalable, vector figure — the best choice for
+    publication.
+-   **Requirements**: every plot is an interactive htmlwidget, so export
+    drives each one with a headless Chrome via the `webshot2` and
+    `chromote` packages, then reformats with `magick`. Install Google
+    Chrome or Chromium if you don’t already have one; the function stops
+    early with a clear message if no browser is found.
+
 ## Usage
 
 ### Dashboard navigation
diff --git a/man/exportAMRVisualizations.Rd b/man/exportAMRVisualizations.Rd
new file mode 100644
index 0000000..595f1cb
--- /dev/null
+++ b/man/exportAMRVisualizations.Rd
@@ -0,0 +1,83 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/export.R
+\name{exportAMRVisualizations}
+\alias{exportAMRVisualizations}
+\title{Export all amRviz dashboard visualizations to static image files}
+\usage{
+exportAMRVisualizations(
+  output_dir = "amRviz_exports",
+  formats = c("png", "pdf"),
+  results_root = NULL,
+  amrdata_root = NULL,
+  species = NULL,
+  top_n_features = 10,
+  network_top_n = 5,
+  width = 1200,
+  height = 800,
+  scale = 2,
+  delay = 1.5,
+  verbose = TRUE
+)
+}
+\arguments{
+\item{output_dir}{Directory to write figures into; created if needed. Files
+are organised as \verb{output_dir/<species>/<panel>.<ext>}, with global
+overviews under \verb{output_dir/_overview} and \verb{output_dir/_across_species}.}
+
+\item{formats}{Any of \code{"png"}, \code{"pdf"}, \code{"jpg"}, \code{"svg"}.}
+
+\item{results_root}{Directory of amRml model outputs (per-species subdirs
+of \verb{*_perf.parquet} / \verb{*_top_features.parquet} / \code{metadata.parquet}).
+\code{NULL} uses the packaged demo data.}
+
+\item{amrdata_root}{Directory of amRdata annotation parquets (for COG
+enrichment). \code{NULL} tries \verb{~/amRdata/data}, else falls back to unenriched.}
+
+\item{species}{Optional character vector restricting which species folders
+to export.}
+
+\item{top_n_features, network_top_n}{Feature counts for the
+feature-importance panels and the drug-feature network.}
+
+\item{width, height}{Viewport size in pixels.}
+
+\item{scale}{Device-pixel multiplier for the PNG render; higher = sharper.}
+
+\item{delay}{Seconds to wait for each widget's JavaScript to settle.}
+
+\item{verbose}{Per-figure progress messages.}
+}
+\value{
+Invisibly, a data frame with one row per figure: \code{group}, \code{name},
+\code{written} (file count), \code{ok}.
+}
+\description{
+Renders every amRviz figure (metadata, model performance, feature
+importance, cross-model holdouts, drug-feature networks) to files, without
+launching the Shiny app. One figure set per species, with global overviews
+under \verb{_overview} / \verb{_across_species}. Selections mirror the dashboard's
+defaults; \code{top_n_features} and \code{network_top_n} override the corresponding
+sliders.
+}
+\details{
+Widgets are driven by a headless Chrome: \code{png} is a webshot2 snapshot;
+\code{jpg} and \code{pdf} are re-encodes of that PNG via \pkg{magick} (identical
+crop and dimensions); \code{svg} is extracted from the DOM via \pkg{chromote},
+using \code{Plotly.toImage} for plotly so titles and labels survive.
+}
+\examples{
+if (interactive()) {
+  # Export the packaged demo figures as PNG + PDF into ./amRviz_exports
+  exportAMRVisualizations()
+
+  # Your own results, PNG only, one species, more features per panel
+  exportAMRVisualizations(
+    output_dir = "figs",
+    formats = "png",
+    results_root = "~/my_amRml_results",
+    species = "Shigella_flexneri",
+    top_n_features = 25,
+    network_top_n = 10
+  )
+}
+}
diff --git a/tests/testthat/test-export.R b/tests/testthat/test-export.R
new file mode 100644
index 0000000..a08c68f
--- /dev/null
+++ b/tests/testthat/test-export.R
@@ -0,0 +1,178 @@
+## Tests for the headless visualization exporter:
+## exportAMRVisualizations() and its internal plan/snapshot helpers.
+
+# ── .exportPlanSpecs (no browser needed) ─────────────────────────────────────
+
+test_that(".exportPlanSpecs builds figure specs from the bundled demo data", {
+  perf <- loadMLResults(verbose = FALSE)
+  top <- loadTopFeat(verbose = FALSE)
+  skip_if(!nrow(perf), "No demo performance data available")
+
+  pairs <- perf |>
+    dplyr::filter(!(species %in% c("cross", "MDR"))) |>
+    dplyr::distinct(species, species_label)
+  ml_species <- list(
+    code = as.character(pairs$species),
+    label = as.character(pairs$species_label)
+  )
+  extdata <- system.file("extdata", package = "amRviz")
+  meta_species <- basename(
+    list.dirs(extdata, full.names = TRUE, recursive = FALSE)
+  )
+  meta_species <- meta_species[vapply(
+    file.path(extdata, meta_species, "metadata.parquet"),
+    file.exists, logical(1)
+  )]
+
+  specs <- .exportPlanSpecs(
+    perf, top, ml_species, meta_species,
+    results_root = NULL, amrdata_root = NULL
+  )
+
+  expect_gt(length(specs), 0)
+  # Every spec is addressable and lazily builds a widget.
+  for (s in specs) {
+    expect_true(all(c("group", "name", "build") %in% names(s)))
+    expect_true(is.function(s$build))
+    expect_true(nzchar(s$group) && nzchar(s$name))
+  }
+  # Global overviews are present exactly once each.
+  groups_names <- vapply(
+    specs, function(s) paste(s$group, s$name), character(1)
+  )
+  expect_true(any(grepl("_overview mcc_strip", groups_names)))
+  expect_true(any(grepl("_overview mcc_heatmap", groups_names)))
+
+  # No duplicate group/name specs: a phantom NA/blank species code shares a
+  # species_label folder with a real one and would otherwise overwrite its
+  # figures (regression guard).
+  expect_equal(anyDuplicated(groups_names), 0L)
+})
+
+test_that("exportAMRVisualizations drops species with no code (no phantom folder)", {
+  perf <- loadMLResults(verbose = FALSE)
+  skip_if(!nrow(perf), "No demo performance data available")
+  # The demo Shigella_sonnei parquets carry some rows with a blank/NA species
+  # code; those must not produce their own export iteration.
+  skip_if(
+    !any(is.na(perf$species) | !nzchar(perf$species)),
+    "Demo data has no blank species codes to exercise the guard"
+  )
+  top <- loadTopFeat(verbose = FALSE)
+  # Deliberately build ml_species the naive way, INCLUDING the phantom
+  # NA/blank-code entry, to prove .exportPlanSpecs drops it defensively.
+  pairs <- perf |>
+    dplyr::filter(!(species %in% c("cross", "MDR"))) |>
+    dplyr::distinct(species, species_label)
+  ml_species <- list(
+    code = as.character(pairs$species),
+    label = as.character(pairs$species_label)
+  )
+  expect_true(any(is.na(ml_species$code) | !nzchar(ml_species$code)))
+
+  specs <- .exportPlanSpecs(
+    perf, top, ml_species, character(0),
+    results_root = NULL, amrdata_root = NULL
+  )
+  groups_names <- vapply(
+    specs, function(s) paste(s$group, s$name), character(1)
+  )
+  expect_equal(anyDuplicated(groups_names), 0L)
+})
+
+# ── caller-adjustable feature counts ─────────────────────────────────────────
+
+test_that("network_top_n flows through .exportPlanSpecs to the network widget", {
+  perf <- loadMLResults(verbose = FALSE)
+  top <- loadTopFeat(verbose = FALSE)
+  skip_if(!nrow(top), "No demo top-feature data available")
+
+  pairs <- perf |>
+    dplyr::filter(!is.na(species) & nzchar(species)) |>
+    dplyr::filter(!(species %in% c("cross", "MDR"))) |>
+    dplyr::distinct(species, species_label)
+  ml_species <- list(
+    code = as.character(pairs$species),
+    label = as.character(pairs$species_label)
+  )
+
+  network_nodes <- function(ntn) {
+    specs <- .exportPlanSpecs(
+      perf, top, ml_species, character(0),
+      results_root = NULL, amrdata_root = NULL, network_top_n = ntn
+    )
+    spec <- Filter(function(s) s$name == "drug_feature_network", specs)[[1]]
+    w <- spec$build()
+    nrow(w$x$nodes)
+  }
+
+  # A larger top-n keeps more features per drug, so the graph has more nodes.
+  expect_gt(network_nodes(12), network_nodes(5))
+})
+
+test_that("exportAMRVisualizations rejects invalid feature counts", {
+  skip_if_not_installed("webshot2")
+  skip_if_not_installed("htmlwidgets")
+  skip_if(
+    tryCatch(!nzchar(chromote::find_chrome()), error = function(e) TRUE),
+    "No Chrome/Chromium available"
+  )
+  expect_error(
+    exportAMRVisualizations(top_n_features = 0, verbose = FALSE),
+    "top_n_features"
+  )
+  expect_error(
+    exportAMRVisualizations(network_top_n = -1, verbose = FALSE),
+    "network_top_n"
+  )
+})
+
+# ── format validation (no browser needed when it errors early) ───────────────
+
+test_that("exportAMRVisualizations rejects unsupported formats", {
+  skip_if_not_installed("webshot2")
+  skip_if_not_installed("htmlwidgets")
+  skip_if(
+    tryCatch(!nzchar(chromote::find_chrome()), error = function(e) TRUE),
+    "No Chrome/Chromium available"
+  )
+  expect_error(
+    exportAMRVisualizations(
+      output_dir = tempfile("amrviz_bad"),
+      formats = "tiff", verbose = FALSE
+    ),
+    "Unsupported format"
+  )
+})
+
+# ── end-to-end export (needs a headless browser) ─────────────────────────────
+
+test_that("exportAMRVisualizations writes files for the demo data", {
+  skip_on_cran()
+  skip_if_not_installed("webshot2")
+  skip_if_not_installed("htmlwidgets")
+  skip_if(
+    tryCatch(!nzchar(chromote::find_chrome()), error = function(e) TRUE),
+    "No Chrome/Chromium available"
+  )
+
+  out <- tempfile("amrviz_export")
+  on.exit(unlink(out, recursive = TRUE), add = TRUE)
+
+  res <- exportAMRVisualizations(
+    output_dir = out,
+    formats = "png",
+    species = "Shigella_flexneri",
+    verbose = FALSE
+  )
+
+  expect_s3_class(res, "data.frame")
+  expect_true(all(c("group", "name", "written", "ok") %in% names(res)))
+  expect_gt(nrow(res), 0)
+  # At least the core per-species figures should render.
+  expect_true(any(res$ok))
+  # No two figures share a group/name (which would mean colliding output paths).
+  expect_equal(anyDuplicated(paste(res$group, res$name)), 0L)
+  png_files <- list.files(out, pattern = "\\.png$", recursive = TRUE)
+  expect_gt(length(png_files), 0)
+})
diff --git a/vignettes/using-amr-dashboard.Rmd b/vignettes/using-amr-dashboard.Rmd
index ec3905e..dfca1ba 100644
--- a/vignettes/using-amr-dashboard.Rmd
+++ b/vignettes/using-amr-dashboard.Rmd
@@ -203,6 +203,67 @@ Browse and export the underlying tables. The **Performance Metrics** and **Top
 Features** tabs render the raw amRml output as searchable tables; add or remove
 columns and download the customized view as CSV for downstream analysis.
 
+## Exporting figures without the dashboard
+
+Sometimes you want the figures, not the interactive app — for a report, a
+manuscript, or an automated pipeline. `exportAMRVisualizations()` renders every
+visualization the dashboard produces to static image files, without launching
+Shiny:
+
+```{r, eval = FALSE}
+# Bundled demo data -> ./amRviz_exports/ as PNG + PDF
+exportAMRVisualizations()
+
+# Your own results, choosing formats, output location, and species
+exportAMRVisualizations(
+  output_dir   = "figures",
+  formats      = c("png", "pdf", "jpg"),
+  results_root = "/path/to/your/amRml/results",
+  species      = c("Shigella_flexneri", "Shigella_sonnei")
+)
+```
+
+The exporter produces one figure set per species, using the same default
+selections each tab opens with (all molecular scales, binary encoding,
+gentamicin where present, the top drug classes for the metadata Sankey, and so
+on). Two of those defaults are adjustable from the call: `top_n_features` sets
+how many features each feature-importance panel shows, and `network_top_n` sets
+how many top features per drug go into the drug-feature network. Files are laid
+out by species:
+
+```
+figures/
+├── _overview/                        # performance heatmaps (all species)
+│   ├── mcc_strip_by_species_scale.png
+│   └── mcc_heatmap_overview.png
+├── _across_species/                  # cross-species feature importance
+│   └── feature_importance_GEN.png
+└── Shigella_flexneri/
+    ├── model_performance.png
+    ├── feature_importance_across_drugs.png
+    ├── cog_categories.png
+    ├── holdout_ridge_country.png
+    ├── drug_feature_network.png
+    ├── metadata_geographic.png
+    └── ...
+```
+
+A few practical notes:
+
+- **Formats.** All four (`png`, `jpg`, `pdf`, `svg`) are supported. `png` is a
+  headless-Chrome screenshot; `jpg` and `pdf` are re-encodes of that PNG (so
+  every raster format has the same crop and dimensions). `svg` is extracted
+  from the rendered DOM and inlines every text element so nothing is lost.
+- **How it works.** Every amRviz plot is an interactive `plotly` or `networkD3`
+  htmlwidget, so there is no server-side raster renderer. Export instead
+  drives each widget with a headless Chrome through the `webshot2` and
+  `chromote` packages, then reformats via `magick`. Install Google Chrome or
+  Chromium if you do not already have one; `exportAMRVisualizations()` stops
+  early with an informative message if no browser is found.
+- **Return value.** The function invisibly returns a data frame summarizing each
+  attempted figure (its group, name, number of files written, and whether it
+  succeeded), which is useful for logging in batch runs.
+
 ## Session information
 
 ```{r}