diff --git a/star/src/lower/visit.rs b/star/src/lower/visit.rs
index e1fdd40..4b74776 100644
--- a/star/src/lower/visit.rs
+++ b/star/src/lower/visit.rs
@@ -2,6 +2,7 @@ use std::{collections::HashMap, str::FromStr};
 
 use euclid::default::Transform2D;
 use log::{debug, warn};
+use lyon_geom::{Box2D, Point};
 use roxmltree::{Document, Node};
 use svgtypes::{
     AspectRatio, LengthListParser, PathParser, PathSegment, PointsParser, TransformListParser,
@@ -172,6 +173,10 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
     fn visit_enter(&mut self, node: Node) {
         use PathSegment::*;
 
+        let mut view_box_opt = None;
+        let mut viewport_pos_opt = [None, None];
+        let mut viewport_size_opt = [1., 1.];
+
         if node.tag_name().name() == CLIP_PATH_TAG_NAME {
             warn!("Clip paths are not supported: {:?}", node);
         }
@@ -266,6 +271,10 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
 
             let viewport_pos = ["x", "y"].map(|attr| self.length_attr_to_user_units(&node, attr));
 
+            view_box_opt = view_box;
+            viewport_pos_opt = viewport_pos;
+            viewport_size_opt = viewport_size;
+
             self.viewport_dim_stack
                 .push(match (view_box.as_ref(), &viewport_size) {
                     (Some(ViewBox { w, h, .. }), _) => [*w, *h],
@@ -333,6 +342,9 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
                 flattened_transform = flattened_transform.then(&viewport_transform);
                 // Does not need Y-axis translation unlike <svg>, already in g-code coords space.
             }
+            view_box_opt = view_box;
+            viewport_pos_opt = [None, None];
+            viewport_size_opt = viewport_size;
         } else if node.has_attribute("viewBox") {
             warn!("View box is not supported on a {}", node.tag_name().name());
         }
@@ -341,6 +353,21 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
         self.name_stack
             .push(node_name(&node, &self.config.extra_attribute_name));
 
+        if node.has_tag_name(SVG_TAG_NAME) || node.has_tag_name(SYMBOL_TAG_NAME) {
+            let viewport_box = if let Some(vb) = view_box_opt {
+                Box2D::new(Point::new(vb.x, vb.y), Point::new(vb.x + vb.w, vb.y + vb.h))
+            } else {
+                let px = viewport_pos_opt[0].unwrap_or(0.);
+                let py = viewport_pos_opt[1].unwrap_or(0.);
+                Box2D::new(
+                    Point::new(px, py),
+                    Point::new(px + viewport_size_opt[0], py + viewport_size_opt[1]),
+                )
+            };
+
+            self.terrarium.push_viewport_bounds(viewport_box);
+        }
+
         if !self.should_draw_node(node) {
             return;
         }
@@ -635,8 +662,33 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
                     .length_attr_to_user_units(&node, "height")
                     .unwrap_or(0.);
 
+                let preserve_aspect_ratio = node
+                    .attribute("preserveAspectRatio")
+                    .map(|attr| {
+                        AspectRatio::from_str(attr).expect("could not parse preserveAspectRatio")
+                    })
+                    .unwrap_or(svgtypes::AspectRatio {
+                        defer: false,
+                        align: svgtypes::Align::XMidYMid,
+                        slice: false,
+                    });
+
+                let view_box = svgtypes::ViewBox {
+                    x: 0.,
+                    y: 0.,
+                    w: image.width() as f64,
+                    h: image.height() as f64,
+                };
+                let image_to_user = get_viewport_transform(
+                    view_box,
+                    Some(preserve_aspect_ratio),
+                    [width, height],
+                    [Some(x), Some(y)],
+                );
+
                 self.comment();
-                self.terrarium.image(image, x, y, width, height);
+                self.terrarium
+                    .image(image, image_to_user, preserve_aspect_ratio);
             }
             // No-op tags
             SVG_TAG_NAME | GROUP_TAG_NAME | USE_TAG_NAME | SYMBOL_TAG_NAME => {}
@@ -651,6 +703,7 @@ impl<'a, 'input, T: Turtle> XmlVisitor for ConversionVisitor<'a, 'input, T> {
         self.name_stack.pop();
         if matches!(node.tag_name().name(), SVG_TAG_NAME | SYMBOL_TAG_NAME) {
             self.viewport_dim_stack.pop();
+            self.terrarium.pop_bounds();
         }
     }
 }
diff --git a/star/src/turtle/elements/image_ops.rs b/star/src/turtle/elements/image_ops.rs
new file mode 100644
index 0000000..6f498e6
--- /dev/null
+++ b/star/src/turtle/elements/image_ops.rs
@@ -0,0 +1,267 @@
+//! Image transformation and cropping utilities to normalize images before passing them to turtles.
+
+use euclid::default::Transform2D;
+use image::{DynamicImage, GenericImage, GenericImageView};
+use lyon_geom::{Box2D, point, vector};
+use svgtypes::AspectRatio;
+
+use super::RasterImage;
+
+/// Transforms (rotates, scales, resamples) the input image using the given transform and aspect ratio.
+/// Returns the transformed image and its bounding box in final coordinate space.
+///
+/// # Note on Scaling
+/// For memory reasons, the pixel dimensions of the returned `DynamicImage` likely won't map directly to the
+/// physical dimensions of the returned `Box2D`.
+///
+/// - axis-aligned transforms defer scaling to `Turtle`-time, where a more efficient operation could be applied
+/// - non-aligned transforms do require resizing, but this is clamped to avoid more than 4x memory usage
+///
+/// The `Turtle` must handle any stretching/scaling as appropriate.
+pub fn transform_image(
+    mut image: DynamicImage,
+    image_to_user: Transform2D<f64>,
+    user_to_final: &Transform2D<f64>,
+    preserve_aspect_ratio: AspectRatio,
+) -> (DynamicImage, Box2D<f64>) {
+    // TODO: should this be more coarse?
+    const EPSILON: f64 = f64::EPSILON;
+    const MAX_NON_AFFINE_SCALE: f64 = 2.;
+
+    let orig_w = image.width();
+    let orig_h = image.height();
+
+    let pixel_bounds = Box2D::new(point(0., 0.), point(orig_w as f64, orig_h as f64));
+    let user_bounds = image_to_user.outer_transformed_box(&pixel_bounds);
+    let image_to_final = image_to_user.then(user_to_final);
+    let transformed_box = image_to_final.outer_transformed_box(&pixel_bounds);
+
+    let (is_transform_axis_aligned, [transformed_x_axis, transformed_y_axis]) = {
+        let tx = user_to_final.transform_vector(vector(1.0, 0.0));
+        let ty = user_to_final.transform_vector(vector(0.0, 1.0));
+
+        let aligned = (tx.y.abs() < EPSILON && ty.x.abs() < EPSILON)
+            || (tx.x.abs() < EPSILON && ty.y.abs() < EPSILON);
+
+        (aligned, [tx, ty])
+    };
+
+    let aspect_ratios_match = {
+        let orig_aspect_ratio = orig_w as f64 / orig_h as f64;
+        let final_aspect_ratio = user_bounds.width() / user_bounds.height();
+        (orig_aspect_ratio - final_aspect_ratio).abs() < EPSILON
+    };
+    let is_simple_orthogonal_rotation = is_transform_axis_aligned
+        && (preserve_aspect_ratio.align == svgtypes::Align::None || aspect_ratios_match);
+
+    if is_simple_orthogonal_rotation {
+        if transformed_x_axis.x.abs() < EPSILON && transformed_y_axis.y.abs() < EPSILON {
+            if transformed_x_axis.y > 0.0 && transformed_y_axis.x < 0.0 {
+                image = image.rotate90();
+            } else if transformed_x_axis.y < 0.0 && transformed_y_axis.x > 0.0 {
+                image = image.rotate270();
+            }
+        } else if transformed_x_axis.y.abs() < EPSILON
+            && transformed_y_axis.x.abs() < EPSILON
+            && transformed_x_axis.x < 0.0
+            && transformed_y_axis.y < 0.0
+        {
+            image = image.rotate180();
+        }
+    } else {
+        // During non-aligned rotation, the corners need to be transparent
+        image = add_alpha_channel(image);
+
+
+
+        let scale_x = image_to_final.transform_vector(vector(1.0, 0.0)).length();
+        let scale_y = image_to_final.transform_vector(vector(0.0, 1.0)).length();
+
+        // Clamp the scale factors proportionally (preserve aspect ratio) to avoid an unnecessarily
+        // large buffer.
+        let max_scale = scale_x.max(scale_y);
+        let (resample_scale_x, resample_scale_y) = if max_scale > MAX_NON_AFFINE_SCALE {
+            (
+                scale_x * (MAX_NON_AFFINE_SCALE / max_scale),
+                scale_y * (MAX_NON_AFFINE_SCALE / max_scale),
+            )
+        } else {
+            (scale_x, scale_y)
+        };
+
+        let new_w = (orig_w as f64 * resample_scale_x).ceil() as u32;
+        let new_h = (orig_h as f64 * resample_scale_y).ceil() as u32;
+
+        let mut new_img = DynamicImage::new(new_w, new_h, image.color());
+
+        if let Some(inv) = image_to_final.inverse() {
+            let min = transformed_box.min;
+            let tb_w = transformed_box.width();
+            let tb_h = transformed_box.height();
+            for new_y in 0..new_h {
+                for new_x in 0..new_w {
+                    // Map new pixel to target space
+                    let target_pt = min
+                        + vector(
+                            (new_x as f64 / new_w as f64) * tb_w,
+                            (new_y as f64 / new_h as f64) * tb_h,
+                        );
+
+                    let orig_pt = inv.transform_point(target_pt);
+                    if (0.0..orig_w as f64).contains(&orig_pt.x)
+                        && (0.0..orig_h as f64).contains(&orig_pt.y)
+                        && let Some(pixel) =
+                            sample_lanczos3(&image, orig_pt.x, orig_pt.y, orig_w, orig_h)
+                    {
+                        new_img.put_pixel(new_x, new_y, pixel);
+                    }
+                }
+            }
+            image = new_img;
+        }
+    }
+
+    (image, transformed_box)
+}
+
+/// Crops the image bounds to the specified viewport bounds, returning None if the cropped area is empty.
+pub fn crop_image_to_bounds(
+    mut image: DynamicImage,
+    img_bbox: Box2D<f64>,
+    bounds: Box2D<f64>,
+) -> Option<RasterImage> {
+    let cropped_bbox = img_bbox.intersection(&bounds)?;
+    if cropped_bbox.is_empty() || img_bbox.is_empty() {
+        return None;
+    }
+
+    let src_w = image.width();
+    let src_h = image.height();
+    if src_w == 0 || src_h == 0 {
+        return None;
+    }
+
+    let orig_w = img_bbox.width();
+    let orig_h = img_bbox.height();
+    let w_crop = cropped_bbox.width();
+    let h_crop = cropped_bbox.height();
+
+    let crop_offset = cropped_bbox.min - img_bbox.min;
+
+    let mut pixel_x = (crop_offset.x / orig_w * src_w as f64).round() as u32;
+    let mut pixel_y = (crop_offset.y / orig_h * src_h as f64).round() as u32;
+
+    pixel_x = pixel_x.min(src_w - 1);
+    pixel_y = pixel_y.min(src_h - 1);
+
+    let pixel_w = ((w_crop / orig_w) * src_w as f64).round() as u32;
+    let pixel_h = ((h_crop / orig_h) * src_h as f64).round() as u32;
+
+    let pixel_w = pixel_w.clamp(1, src_w - pixel_x);
+    let pixel_h = pixel_h.clamp(1, src_h - pixel_y);
+
+    image = image.crop_imm(pixel_x, pixel_y, pixel_w, pixel_h);
+
+    Some(RasterImage {
+        dimensions: cropped_bbox,
+        image,
+    })
+}
+
+fn add_alpha_channel(image: DynamicImage) -> DynamicImage {
+    if image.color().has_alpha() {
+        image
+    } else {
+        match image {
+            DynamicImage::ImageLuma8(_) => DynamicImage::ImageLumaA8(image.to_luma_alpha8()),
+            DynamicImage::ImageLuma16(_) => DynamicImage::ImageLumaA16(image.to_luma_alpha16()),
+            DynamicImage::ImageRgb8(_) => DynamicImage::ImageRgba8(image.to_rgba8()),
+            DynamicImage::ImageRgb16(_) => DynamicImage::ImageRgba16(image.to_rgba16()),
+            DynamicImage::ImageRgb32F(_) => DynamicImage::ImageRgba32F(image.to_rgba32f()),
+            other => DynamicImage::ImageRgba8(other.to_rgba8()),
+        }
+    }
+}
+
+/// <https://en.wikipedia.org/wiki/Lanczos_resampling#Lanczos_kernel>
+fn lanczos3_weight(x: f64) -> f64 {
+    const A: f64 = 3.;
+
+    let x = x.abs();
+    if x < f64::EPSILON {
+        1.0
+    } else if x < A {
+        // sinc(x)*sinc(x/a)
+        let pi_x = std::f64::consts::PI * x;
+        (pi_x.sin() * (pi_x / A).sin()) / (pi_x * pi_x / A)
+    } else {
+        // Does not contribute
+        0.0
+    }
+}
+
+/// TODO: Bad attempt at lanczos3 resample for non-aligned transforms that should probably be revisited...
+fn sample_lanczos3(
+    image: &image::DynamicImage,
+    x0: f64,
+    y0: f64,
+    orig_w: u32,
+    orig_h: u32,
+) -> Option<image::Rgba<u8>> {
+    let x_floor = x0.floor() as i64;
+    let y_floor = y0.floor() as i64;
+
+    // Pre-compute horizontal (x) and vertical (y) weights in 1D arrays
+    let mut wx_arr = [0.0; 6];
+    for (dx, wx) in wx_arr.iter_mut().enumerate() {
+        let px = x_floor - 2 + dx as i64;
+        *wx = lanczos3_weight(x0 - px as f64);
+    }
+
+    let mut wy_arr = [0.0; 6];
+    for (dy, wy) in wy_arr.iter_mut().enumerate() {
+        let py = y_floor - 2 + dy as i64;
+        *wy = lanczos3_weight(y0 - py as f64);
+    }
+
+    let mut r_sum = 0.0;
+    let mut g_sum = 0.0;
+    let mut b_sum = 0.0;
+    let mut a_sum = 0.0;
+    let mut w_sum = 0.0;
+
+    for (dy_idx, &wy) in wy_arr.iter().enumerate() {
+        if wy.abs() < f64::EPSILON {
+            continue;
+        }
+        let py = y_floor - 2 + dy_idx as i64;
+        let clamped_py = py.clamp(0, orig_h as i64 - 1) as u32;
+
+        for (dx_idx, &wx) in wx_arr.iter().enumerate() {
+            let w = wx * wy;
+            if w.abs() < f64::EPSILON {
+                continue;
+            }
+            let px = x_floor - 2 + dx_idx as i64;
+            let clamped_px = px.clamp(0, orig_w as i64 - 1) as u32;
+
+            let pixel = image.get_pixel(clamped_px, clamped_py);
+            r_sum += pixel.0[0] as f64 * w;
+            g_sum += pixel.0[1] as f64 * w;
+            b_sum += pixel.0[2] as f64 * w;
+            a_sum += pixel.0[3] as f64 * w;
+            w_sum += w;
+        }
+    }
+
+    if w_sum > 0.0 {
+        let r = (r_sum / w_sum).clamp(0.0, 255.0).round() as u8;
+        let g = (g_sum / w_sum).clamp(0.0, 255.0).round() as u8;
+        let b = (b_sum / w_sum).clamp(0.0, 255.0).round() as u8;
+        let a = (a_sum / w_sum).clamp(0.0, 255.0).round() as u8;
+
+        Some(image::Rgba([r, g, b, a]))
+    } else {
+        None
+    }
+}
diff --git a/star/src/turtle/elements/mod.rs b/star/src/turtle/elements/mod.rs
index ecab444..9e4f3be 100644
--- a/star/src/turtle/elements/mod.rs
+++ b/star/src/turtle/elements/mod.rs
@@ -18,6 +18,9 @@ mod tsp;
 
 pub(crate) mod fill;
 
+#[cfg(feature = "image")]
+pub(crate) mod image_ops;
+
 /// Defines the algorithm used to calculate how a polygon is filled.
 ///
 /// <https://www.w3.org/TR/SVG/painting.html#FillRuleProperty>
diff --git a/star/src/turtle/mod.rs b/star/src/turtle/mod.rs
index a21507e..d7636a7 100644
--- a/star/src/turtle/mod.rs
+++ b/star/src/turtle/mod.rs
@@ -1,7 +1,7 @@
 use std::fmt::Debug;
 
 use lyon_geom::{
-    ArcFlags, CubicBezierSegment, Point, QuadraticBezierSegment, SvgArc, Vector,
+    ArcFlags, Box2D, CubicBezierSegment, Point, QuadraticBezierSegment, SvgArc, Vector,
     euclid::{Angle, default::Transform2D},
     point, vector,
 };
@@ -62,6 +62,8 @@ pub(crate) struct Terrarium<T: Turtle + std::fmt::Debug> {
     previous_quadratic_control: Option<Point<f64>>,
     previous_cubic_control: Option<Point<f64>>,
     comment: Option<String>,
+    current_bounds: Option<Box2D<f64>>,
+    bounds_stack: Vec<Option<Box2D<f64>>>,
 }
 
 impl<T: Turtle + std::fmt::Debug> Terrarium<T> {
@@ -77,9 +79,29 @@ impl<T: Turtle + std::fmt::Debug> Terrarium<T> {
             previous_quadratic_control: None,
             previous_cubic_control: None,
             comment: None,
+            current_bounds: None,
+            bounds_stack: vec![],
         }
     }
 
+    pub fn push_viewport_bounds(&mut self, viewport_box: Box2D<f64>) {
+        let new_bounds = self.current_transform.outer_transformed_box(&viewport_box);
+        let combined_bounds = match self.current_bounds {
+            Some(parent) => parent.intersection(&new_bounds),
+            None => Some(new_bounds),
+        };
+        self.push_bounds(combined_bounds);
+    }
+
+    pub fn push_bounds(&mut self, bounds: Option<Box2D<f64>>) {
+        self.bounds_stack.push(self.current_bounds);
+        self.current_bounds = bounds;
+    }
+
+    pub fn pop_bounds(&mut self) {
+        self.current_bounds = self.bounds_stack.pop().expect("bounds stack underflow");
+    }
+
     /// Move the turtle to the given absolute/relative coordinates in the current transform
     /// <https://www.w3.org/TR/SVG/paths.html#PathDataMovetoCommands>
     fn move_to(&mut self, abs: bool, x: f64, y: f64) -> Point<f64> {
@@ -371,20 +393,31 @@ impl<T: Turtle + std::fmt::Debug> Terrarium<T> {
 
     /// <https://www.w3.org/TR/SVG/embedded.html#ImageElement>
     #[cfg(feature = "image")]
-    pub fn image(&mut self, image: image::DynamicImage, x: f64, y: f64, width: f64, height: f64) {
-        // Transform the corners to get the final x, y, width, height.
-        let t0 = self.current_transform.transform_point(point(x, y));
-        let t1 = self
-            .current_transform
-            .transform_point(point(x, y) + vector(width, height));
-        self.turtle.image(crate::turtle::elements::RasterImage {
-            // After transformation, the corners may be swapped resulting in a new x y.
-            dimensions: lyon_geom::Box2D::new(
-                point(t0.x.min(t1.x), t0.y.min(t1.y)),
-                point(t0.x.max(t1.x), t0.y.max(t1.y)),
-            ),
+    pub fn image(
+        &mut self,
+        image: image::DynamicImage,
+        image_to_user: Transform2D<f64>,
+        preserve_aspect_ratio: svgtypes::AspectRatio,
+    ) {
+        let (image, transformed_box) = self::elements::image_ops::transform_image(
             image,
-        });
+            image_to_user,
+            &self.current_transform,
+            preserve_aspect_ratio,
+        );
+
+        let final_raster_image = if let Some(bounds) = self.current_bounds {
+            self::elements::image_ops::crop_image_to_bounds(image, transformed_box, bounds)
+        } else {
+            Some(crate::turtle::elements::RasterImage {
+                dimensions: transformed_box,
+                image,
+            })
+        };
+
+        if let Some(raster_image) = final_raster_image {
+            self.turtle.image(raster_image);
+        }
     }
 
     /// Push a generic transform onto the stack
@@ -536,6 +569,8 @@ impl<T: Turtle + std::fmt::Debug> Terrarium<T> {
             previous_quadratic_control: self.previous_quadratic_control,
             previous_cubic_control: self.previous_cubic_control,
             comment: self.comment.clone(),
+            current_bounds: self.current_bounds,
+            bounds_stack: vec![],
         };
         sub.apply_path(segments);
         let segments = sub.finish().into_strokes();