diff --git a/bench/Cargo.toml b/bench/Cargo.toml
index 25a112a16..36f40b8c9 100644
--- a/bench/Cargo.toml
+++ b/bench/Cargo.toml
@@ -5,10 +5,10 @@ authors = ["The Typst Project Developers"]
 edition = "2018"
 publish = false
 
-[dependencies]
+[dev-dependencies]
+criterion = "0.3"
 fontdock = { path = "../../fontdock" }
 typst = { path = ".." }
-criterion = "0.3"
 
 [[bench]]
 name = "typst"
diff --git a/src/eval/mod.rs b/src/eval/mod.rs
index 6d8f66ca6..bab93a058 100644
--- a/src/eval/mod.rs
+++ b/src/eval/mod.rs
@@ -25,7 +25,7 @@ use crate::diag::{Deco, Feedback, Pass};
 use crate::env::SharedEnv;
 use crate::geom::{BoxAlign, Dir, Flow, Gen, Length, Linear, Relative, Sides, Size};
 use crate::layout::{
-    Document, Expansion, LayoutNode, Pad, Pages, Par, Softness, Spacing, Stack, Text,
+    Document, Expansion, LayoutNode, Pad, Pages, Par, Spacing, Stack, Text,
 };
 use crate::syntax::*;
 
@@ -35,9 +35,9 @@ use crate::syntax::*;
 /// evaluation.
 pub fn eval(tree: &SynTree, env: SharedEnv, state: State) -> Pass<Document> {
     let mut ctx = EvalContext::new(env, state);
-    ctx.start_page_group(false);
+    ctx.start_page_group(Softness::Hard);
     tree.eval(&mut ctx);
-    ctx.end_page_group(true);
+    ctx.end_page_group(|s| s == Softness::Hard);
     ctx.finish()
 }
 
@@ -117,28 +117,35 @@ impl EvalContext {
 
     /// Start a page group based on the active page state.
     ///
-    /// If both this `hard` and the one in the matching call to `end_page_group`
-    /// are false, empty page runs will be omitted from the output.
+    /// The `softness` is a hint on whether empty pages should be kept in the
+    /// output.
     ///
     /// This also starts an inner paragraph.
-    pub fn start_page_group(&mut self, hard: bool) {
+    pub fn start_page_group(&mut self, softness: Softness) {
         self.start_group(PageGroup {
             size: self.state.page.size,
             padding: self.state.page.margins(),
             flow: self.state.flow,
             align: self.state.align,
-            hard,
+            softness,
         });
         self.start_par_group();
     }
 
-    /// End a page group and push it to the finished page runs.
+    /// End a page group, returning its [`Softness`].
+    ///
+    /// Whether the page is kept when it's empty is decided by `keep_empty`
+    /// based on its softness. If kept, the page is pushed to the finished page
+    /// runs.
     ///
     /// This also ends an inner paragraph.
-    pub fn end_page_group(&mut self, hard: bool) {
+    pub fn end_page_group(
+        &mut self,
+        keep_empty: impl FnOnce(Softness) -> bool,
+    ) -> Softness {
         self.end_par_group();
         let (group, children) = self.end_group::<PageGroup>();
-        if hard || group.hard || !children.is_empty() {
+        if !children.is_empty() || keep_empty(group.softness) {
             self.runs.push(Pages {
                 size: group.size,
                 child: LayoutNode::dynamic(Pad {
@@ -152,6 +159,7 @@ impl EvalContext {
                 }),
             })
         }
+        group.softness
     }
 
     /// Start a content group.
@@ -273,7 +281,7 @@ struct PageGroup {
     padding: Sides<Linear>,
     flow: Flow,
     align: BoxAlign,
-    hard: bool,
+    softness: Softness,
 }
 
 /// A group for generic content.
@@ -286,6 +294,15 @@ struct ParGroup {
     line_spacing: Length,
 }
 
+/// Defines how items interact with surrounding items.
+#[derive(Debug, Copy, Clone, Eq, PartialEq, Ord, PartialOrd)]
+pub enum Softness {
+    /// Soft items can be skipped in some circumstances.
+    Soft,
+    /// Hard items are always retained.
+    Hard,
+}
+
 /// Evaluate an item.
 ///
 /// _Note_: Evaluation is not necessarily pure, it may change the active state.
diff --git a/src/layout/spacing.rs b/src/layout/spacing.rs
index 1ba3b54a2..c9a9c233e 100644
--- a/src/layout/spacing.rs
+++ b/src/layout/spacing.rs
@@ -1,14 +1,21 @@
 use std::fmt::{self, Debug, Formatter};
 
 use super::*;
+use crate::eval::Softness;
 
 /// A  spacing node.
 #[derive(Copy, Clone, PartialEq)]
 pub struct Spacing {
     /// The amount of spacing to insert.
     pub amount: Length,
-    /// Spacing interaction, see [`Softness`]'s documentation for more
-    /// information.
+    /// Defines how spacing interacts with surrounding spacing.
+    ///
+    /// Hard spacing assures that a fixed amount of spacing will always be
+    /// inserted. Soft spacing will be consumed by previous soft spacing or
+    /// neighbouring hard spacing and can be used to insert overridable spacing,
+    /// e.g. between words or paragraphs.
+    ///
+    /// This field is only used in evaluation, not in layouting.
     pub softness: Softness,
 }
 
@@ -32,18 +39,3 @@ impl From<Spacing> for LayoutNode {
         Self::Spacing(spacing)
     }
 }
-
-/// Defines how spacing interacts with surrounding spacing.
-///
-/// Hard spacing assures that a fixed amount of spacing will always be inserted.
-/// Soft spacing will be consumed by previous soft spacing or neighbouring hard
-/// spacing and can be used to insert overridable spacing, e.g. between words or
-/// paragraphs.
-#[derive(Debug, Copy, Clone, Eq, PartialEq, Ord, PartialOrd)]
-pub enum Softness {
-    /// Soft spacing is not laid out if it directly follows other soft spacing
-    /// or if it touches hard spacing.
-    Soft,
-    /// Hard spacing is always laid out and consumes surrounding soft spacing.
-    Hard,
-}
diff --git a/src/library/layout.rs b/src/library/layout.rs
index 4a787f1ec..36a208219 100644
--- a/src/library/layout.rs
+++ b/src/library/layout.rs
@@ -1,7 +1,8 @@
 use std::fmt::{self, Display, Formatter};
 
+use crate::eval::Softness;
 use crate::geom::{Length, Linear};
-use crate::layout::{Expansion, Fixed, Softness, Spacing, Stack};
+use crate::layout::{Expansion, Fixed, Spacing, Stack};
 use crate::paper::{Paper, PaperClass};
 use crate::prelude::*;
 
@@ -184,8 +185,8 @@ pub fn boxed(mut args: Args, ctx: &mut EvalContext) -> Value {
     let body = args.find::<SynTree>().unwrap_or_default();
     let width = args.get::<_, Linear>(ctx, "width");
     let height = args.get::<_, Linear>(ctx, "height");
-    let main = args.get::<_, Spanned<Dir>>(ctx, "main");
-    let cross = args.get::<_, Spanned<Dir>>(ctx, "cross");
+    let main = args.get::<_, Spanned<Dir>>(ctx, "main-dir");
+    let cross = args.get::<_, Spanned<Dir>>(ctx, "cross-dir");
     ctx.set_flow(Gen::new(main, cross));
     args.done(ctx);
 
@@ -269,7 +270,7 @@ pub fn page(mut args: Args, ctx: &mut EvalContext) -> Value {
     let snapshot = ctx.state.clone();
     let body = args.find::<SynTree>();
 
-    if let Some(paper) = args.find::<Paper>() {
+    if let Some(paper) = args.get::<_, Paper>(ctx, 0) {
         ctx.state.page.class = paper.class;
         ctx.state.page.size = paper.size();
     }
@@ -309,21 +310,24 @@ pub fn page(mut args: Args, ctx: &mut EvalContext) -> Value {
         std::mem::swap(&mut size.width, &mut size.height);
     }
 
-    let main = args.get::<_, Spanned<Dir>>(ctx, "main");
-    let cross = args.get::<_, Spanned<Dir>>(ctx, "cross");
+    let main = args.get::<_, Spanned<Dir>>(ctx, "main-dir");
+    let cross = args.get::<_, Spanned<Dir>>(ctx, "cross-dir");
     ctx.set_flow(Gen::new(main, cross));
 
     args.done(ctx);
 
+    let mut softness = ctx.end_page_group(|_| false);
+
     if let Some(body) = body {
-        ctx.end_page_group(false);
-        ctx.start_page_group(true);
+        // TODO: Restrict body to a single page?
+        ctx.start_page_group(Softness::Hard);
         body.eval(ctx);
+        ctx.end_page_group(|s| s == Softness::Hard);
         ctx.state = snapshot;
+        softness = Softness::Soft;
     }
 
-    ctx.end_page_group(false);
-    ctx.start_page_group(false);
+    ctx.start_page_group(softness);
 
     Value::None
 }
@@ -331,7 +335,7 @@ pub fn page(mut args: Args, ctx: &mut EvalContext) -> Value {
 /// `pagebreak`: Start a new page.
 pub fn pagebreak(args: Args, ctx: &mut EvalContext) -> Value {
     args.done(ctx);
-    ctx.end_page_group(false);
-    ctx.start_page_group(true);
+    ctx.end_page_group(|_| true);
+    ctx.start_page_group(Softness::Hard);
     Value::None
 }
diff --git a/tests/ref/image-error.png b/tests/ref/image-error.png
deleted file mode 100644
index 812a3758c..000000000
Binary files a/tests/ref/image-error.png and /dev/null differ
diff --git a/tests/ref/image-formats.png b/tests/ref/image-formats.png
new file mode 100644
index 000000000..45b23276b
Binary files /dev/null and b/tests/ref/image-formats.png differ
diff --git a/tests/ref/image-jpeg.png b/tests/ref/image-jpeg.png
deleted file mode 100644
index ef9e74cbd..000000000
Binary files a/tests/ref/image-jpeg.png and /dev/null differ
diff --git a/tests/ref/image-png.png b/tests/ref/image-png.png
deleted file mode 100644
index 4e0818d21..000000000
Binary files a/tests/ref/image-png.png and /dev/null differ
diff --git a/tests/ref/page-body.png b/tests/ref/page-body.png
new file mode 100644
index 000000000..a3b3d7264
Binary files /dev/null and b/tests/ref/page-body.png differ
diff --git a/tests/ref/page-dirs.png b/tests/ref/page-dirs.png
new file mode 100644
index 000000000..2bf23ebcc
Binary files /dev/null and b/tests/ref/page-dirs.png differ
diff --git a/tests/ref/page-metrics.png b/tests/ref/page-metrics.png
new file mode 100644
index 000000000..2e8d626db
Binary files /dev/null and b/tests/ref/page-metrics.png differ
diff --git a/tests/typ/example-coma.typ b/tests/typ/example-coma.typ
index f841a1221..d41032c1f 100644
--- a/tests/typ/example-coma.typ
+++ b/tests/typ/example-coma.typ
@@ -1,4 +1,4 @@
-// Small integration test of syntax, page setup, box layout and alignment.
+// Test integration of syntax, page setup, box layout and alignment.
 
 [page: width=450pt, height=300pt, margins=1cm]
 
diff --git a/tests/typ/image-error.typ b/tests/typ/image-error.typ
index 4fde4ab29..9a7f2c400 100644
--- a/tests/typ/image-error.typ
+++ b/tests/typ/image-error.typ
@@ -1,5 +1,8 @@
-// error: 5:9-5:30 failed to load image
+// Test error cases of the `image` function.
+
+// compare-ref: false
 // error: 8:9-8:30 failed to load image
+// error: 11:9-11:30 failed to load image
 
 // File does not exist.
 [image: "path/does/not/exist"]
diff --git a/tests/typ/image-fit.typ b/tests/typ/image-fit.typ
index b735f058e..eca7c1e45 100644
--- a/tests/typ/image-fit.typ
+++ b/tests/typ/image-fit.typ
@@ -1,3 +1,5 @@
+// Test configuring the size and fitting behaviour of images.
+
 // Fit to width of page.
 [image: "res/rhino.png"]
 
diff --git a/tests/typ/image-formats.typ b/tests/typ/image-formats.typ
new file mode 100644
index 000000000..36b991d01
--- /dev/null
+++ b/tests/typ/image-formats.typ
@@ -0,0 +1,8 @@
+// Test loading different image formats.
+
+// Load an RGBA PNG image.
+[image: "res/rhino.png"]
+[pagebreak]
+
+// Load an RGB JPEG image.
+[image: "res/tiger.jpg"]
diff --git a/tests/typ/image-jpeg.typ b/tests/typ/image-jpeg.typ
deleted file mode 100644
index 48cf1a0dc..000000000
--- a/tests/typ/image-jpeg.typ
+++ /dev/null
@@ -1,2 +0,0 @@
-// Load an RGB JPEG image.
-[image: "res/tiger.jpg"]
diff --git a/tests/typ/image-png.typ b/tests/typ/image-png.typ
deleted file mode 100644
index 482591e92..000000000
--- a/tests/typ/image-png.typ
+++ /dev/null
@@ -1,2 +0,0 @@
-// Load an RGBA PNG image.
-[image: "res/rhino.png"]
diff --git a/tests/typ/page-body.typ b/tests/typ/page-body.typ
new file mode 100644
index 000000000..6067dcfee
--- /dev/null
+++ b/tests/typ/page-body.typ
@@ -0,0 +1,11 @@
+// Test a combination of pages with bodies and normal content.
+
+[page: height=50pt]
+
+[page][First]
+[page][Second]
+[pagebreak]
+Fourth
+[page][Fifth]
+Sixth
+[page][Seventh and last]
diff --git a/tests/typ/page-dirs.typ b/tests/typ/page-dirs.typ
new file mode 100644
index 000000000..09ba4e9e8
--- /dev/null
+++ b/tests/typ/page-dirs.typ
@@ -0,0 +1,5 @@
+// Test changing the layouting directions of pages.
+
+[page: main-dir=btt, cross-dir=rtl]
+
+Right to left!
diff --git a/tests/typ/page-error.typ b/tests/typ/page-error.typ
new file mode 100644
index 000000000..dee5c449d
--- /dev/null
+++ b/tests/typ/page-error.typ
@@ -0,0 +1,11 @@
+// Test error cases of the `page` function.
+
+// compare-ref: false
+// error: 8:8-8:19 invalid paper
+// error: 11:17-11:20 aligned axis
+
+// Invalid paper.
+[page: nonexistant]
+
+// Aligned axes.
+[page: main-dir=ltr]
diff --git a/tests/typ/page-metrics.typ b/tests/typ/page-metrics.typ
new file mode 100644
index 000000000..6c7bd4611
--- /dev/null
+++ b/tests/typ/page-metrics.typ
@@ -0,0 +1,25 @@
+// Test configuring page sizes and margins.
+
+// Set width.
+[page: width=50pt][High]
+
+// Set height.
+[page: height=50pt][Wide]
+
+// Set all margins at once.
+[page: margins=40pt][
+    [align: top, left][TL]
+    [align: bottom, right][BR]
+]
+
+// Set individual margins.
+[page: left=0pt >> align: left][Left]
+[page: right=0pt >> align: right][Right]
+[page: top=0pt >> align: top][Top]
+[page: bottom=0pt >> align: bottom][Bottom]
+
+// Ensure that specific margins override general margins.
+[page: margins=0pt, left=40pt][Overriden]
+
+// Flip the page.
+[page: a10, flip=true][Flipped]
diff --git a/tests/typeset.rs b/tests/typeset.rs
index 012aff637..5f1085858 100644
--- a/tests/typeset.rs
+++ b/tests/typeset.rs
@@ -127,7 +127,7 @@ fn test(
 
     let src = fs::read_to_string(src_path).unwrap();
     let map = LineMap::new(&src);
-    let ref_diags = parse_diags(&src, &map);
+    let (ref_diags, compare_ref) = parse_metadata(&src, &map);
 
     let mut state = State::default();
     state.page.size = Size::uniform(Length::pt(120.0));
@@ -167,14 +167,16 @@ fn test(
         }
     }
 
-    if let Ok(ref_pixmap) = Pixmap::load_png(&ref_path) {
-        if canvas.pixmap != ref_pixmap {
-            println!("  Does not match reference image. ❌");
+    if compare_ref {
+        if let Ok(ref_pixmap) = Pixmap::load_png(&ref_path) {
+            if canvas.pixmap != ref_pixmap {
+                println!("  Does not match reference image. ❌");
+                ok = false;
+            }
+        } else {
+            println!("  Failed to open reference image. ❌");
             ok = false;
         }
-    } else {
-        println!("  Failed to open reference image. ❌");
-        ok = false;
     }
 
     if ok {
@@ -184,10 +186,13 @@ fn test(
     ok
 }
 
-fn parse_diags(src: &str, map: &LineMap) -> SpanVec<Diag> {
+fn parse_metadata(src: &str, map: &LineMap) -> (SpanVec<Diag>, bool) {
     let mut diags = vec![];
+    let mut compare_ref = true;
 
     for line in src.lines() {
+        compare_ref &= !line.starts_with("// compare-ref: false");
+
         let (level, rest) = if let Some(rest) = line.strip_prefix("// error: ") {
             (Level::Error, rest)
         } else if let Some(rest) = line.strip_prefix("// warning: ") {
@@ -211,7 +216,8 @@ fn parse_diags(src: &str, map: &LineMap) -> SpanVec<Diag> {
     }
 
     diags.sort();
-    diags
+
+    (diags, compare_ref)
 }
 
 fn print_diag(diag: &Spanned<Diag>, map: &LineMap) {