DemchaAV · DemchaAV · Jun 3, 2026 · Jun 3, 2026
@@ -17,6 +17,17 @@ Housekeeping cycle plus the public pixel-level visual-regression API (Track N).
   consumers can now run the same render-PDF → diff-PNG baseline gate against
   their own presets and templates instead of copying the harness. Behaviour is
   unchanged; the PDF→image step is inlined on PDFBox's `PDFRenderer`.
+- Exposed `PdfVisualRegression.APPROVE_PROPERTY` (`@since 1.6.9`) — the
+  `graphcompose.visual.approve` system-property name — so consumers can toggle
+  baseline-approve mode without hard-coding the string (mirrors
+  `LayoutSnapshotAssertions.UPDATE_PROPERTY`).
+
+### Documentation
+
+- Added [`docs/operations/visual-regression-testing.md`](docs/operations/visual-regression-testing.md):
+  pixel-vs-semantic guidance, the `PdfVisualRegression` API, approve mode,
+  baseline layout, and cross-platform tolerance calibration.
+- README "Which API should I use?" gains a pixel-level visual-regression row.
 
 ## v1.6.8 — 2026-06-01
 

@@ -59,7 +59,7 @@ Sits between **iText** (low-level page primitives) and **JasperReports** (XML-te
 
 - **Server-side PDF generation in Java** &mdash; invoices, CVs, reports, proposals, statements, schedules.
 - **Templated documents from data** &mdash; themed presets (`ModernProfessional`, `InvoiceTemplateV2`, &hellip;) you parameterise instead of re-styling every time.
-- **Regression-tested layouts** &mdash; `DocumentSession#layoutSnapshot()` makes layout changes visible in PRs before any byte ships.
+- **Regression-tested layouts** &mdash; `DocumentSession#layoutSnapshot()` makes layout changes visible in PRs before any byte ships; `PdfVisualRegression` adds a pixel-level gate for font and colour fidelity.
 - **Streaming PDFs from web backends** &mdash; Spring Boot `@RestController` writing straight to the response ([`HttpStreamingExample`](./examples/src/main/java/com/demcha/examples/features/streaming/HttpStreamingExample.java)).
 - **Higher-level than PDFBox, lighter than JasperReports** &mdash; Java DSL describes semantics; no XML templates, no manual coordinates.
 
@@ -90,6 +90,7 @@ GraphCompose uses PDFBox under the hood as the rendering backend &mdash; the com
 | Generate a CV / cover letter from data | Layered templates | `ModernProfessional.create().compose(session, cvDocument)` &mdash; see [layered templates](./docs/templates/v2-layered/README.md) |
 | Add a custom visual primitive | Engine extension | `NodeDefinition` + `PdfFragmentRenderHandler` &mdash; see [extension guide](./docs/contributing/extension-guide.md) |
 | Regression-test generated layouts | Layout snapshots | `DocumentSession#layoutSnapshot()` &mdash; quickstart at [Testing your document](./docs/operations/test-your-document.md); full reference at [snapshot testing](./docs/operations/layout-snapshot-testing.md) |
+| Pixel-test the rendered PDF (fonts, colours, anti-aliasing) | Visual regression | `PdfVisualRegression.standard()&hellip;assertMatchesBaseline(...)` &mdash; see [visual regression testing](./docs/operations/visual-regression-testing.md) |
 | See the live playground / gallery | Next.js showcase site | [Showcase](https://DemchaAV.github.io/GraphCompose/) &mdash; source under [`site/`](./site), built with `next build` and deployed via the [Pages workflow](./.github/workflows/deploy-site.yml) |
 
 ## Installation

@@ -0,0 +1,139 @@
+# Visual Regression Testing
+
+Pixel-level visual regression is the outermost geometry-fidelity layer in GraphCompose. It renders a PDF to one PNG per page and diffs each page against a committed baseline.
+
+It complements — does not replace — [layout snapshot testing](./layout-snapshot-testing.md):
+
+1. unit tests validate isolated layout math
+2. **layout snapshots** validate the resolved document tree (coordinates, page spans, layer/order) — structural geometry
+3. **visual regression** validates the rendered pixels — font shape, colour, anti-aliasing, glyph fallback
+4. human inspection of the PDF remains the final eye
+
+Reach for visual regression when the failure you care about is *pixel-level* rather than *geometry-level*: the layout snapshot still matches but the PDF looks wrong (wrong font, wrong colour, missing glyph, anti-aliasing drift).
+
+## Pixel vs semantic — which layer?
+
+| You want to catch&hellip; | Use |
+|---|---|
+| A node moved / page break shifted / sibling order changed | layout snapshot (semantic) |
+| The PDF looks identical pixel-for-pixel — fonts, colours, glyphs | **visual regression (pixel)** |
+| A specific layout-math rule | a focused unit test |
+
+The semantic layer is cheap, deterministic, and cross-platform stable. The pixel layer is precise but sensitive to platform font rendering (see [Cross-platform tolerance](#cross-platform-tolerance) below). A flagship template or a preset you publish to others deserves both.
+
+## Public API
+
+The harness is `com.demcha.compose.testing.visual.PdfVisualRegression` (`@since 1.6.9`), a sibling to the semantic `com.demcha.compose.testing.layout.*` helpers. It ships in the main artifact, so library consumers use the exact helpers GraphCompose uses in its own tests.
+
+### Assert a committed baseline
+
+```java
+import com.demcha.compose.GraphCompose;
+import com.demcha.compose.document.api.DocumentPageSize;
+import com.demcha.compose.document.api.DocumentSession;
+import com.demcha.compose.testing.visual.PdfVisualRegression;
+import org.junit.jupiter.api.Test;
+
+class InvoiceVisualParityTest {
+
+    @Test
+    void invoiceRendersPixelIdentical() throws Exception {
+        byte[] pdfBytes;
+        try (DocumentSession document = GraphCompose.document()
+                .pageSize(DocumentPageSize.A4)
+                .margin(22, 22, 22, 22)
+                .create()) {
+            template.compose(document, spec);
+            pdfBytes = document.toPdfBytes();
+        }
+
+        PdfVisualRegression.standard()
+                .assertMatchesBaseline("invoice_standard", pdfBytes);
+    }
+}
+```
+
+`assertMatchesBaseline(name, pdfBytes)` renders every page, compares against `<name>-page-N.png` under the baseline root, and throws `AssertionError` if any page exceeds the configured budget. On failure it writes `<name>-page-N.actual.png` and `<name>-page-N.diff.png` next to the baseline for inspection.
+
+### Configure the harness
+
+`PdfVisualRegression` is immutable; every setter returns a copy.
+
+| Setter | Default | Meaning |
+|---|---|---|
+| `baselineRoot(Path)` | `src/test/resources/visual-baselines` | where baselines and diff sidecars live |
+| `renderScale(float)` | `1.0` | render scale multiplier (`2.0` = retina); must be `> 0` |
+| `perPixelTolerance(int)` | `6` | allowed per-channel delta (`0..255`) before a pixel counts as mismatched |
+| `mismatchedPixelBudget(long)` | `0` | mismatched pixels tolerated per page before the assertion fails |
+
+### Diff images directly
+
+For ad-hoc comparison, render pages and call `ImageDiff` yourself:
+
+```java
+List<BufferedImage> pages = PdfVisualRegression.standard().renderPages(pdfBytes);
+ImageDiff.Result diff = ImageDiff.compare(expectedPng, pages.get(0), 6);
+assertThat(diff.withinBudget(0)).isTrue();
+```
+
+## Approve mode (blessing baselines)
+
+There is no baseline the first time. Run with the approve flag to write the current renders as the baseline:
+
+```bash
+./mvnw test -Dtest=InvoiceVisualParityTest -Dgraphcompose.visual.approve=true
+```
+
+The system-property name is exposed as `PdfVisualRegression.APPROVE_PROPERTY`; the environment variable `GRAPHCOMPOSE_VISUAL_APPROVE=true` works as a fallback. In approve mode the harness writes baselines and skips the diff assertion — so **never enable it in CI verification**, only when you have reviewed the new render and intend to re-bless.
+
+## Where files live
+
+- committed baselines: `<baselineRoot>/<name>-page-N.png`
+- mismatch artifacts (normal runs): `<baselineRoot>/<name>-page-N.actual.png` and `<name>-page-N.diff.png` (mismatched pixels red, matching pixels greyscale)
+
+Use a flat `name`, or pre-create nested baseline directories — the harness creates the baseline root but not intermediate folders.
+
+## Cross-platform tolerance
+
+PDFBox font rasterization drifts slightly across platforms (different system fonts, different rasterizer). A baseline recorded on Windows will not match Linux CI pixel-for-pixel.
+
+The `standard()` defaults are strict (tolerance `6`, budget `0`) — good for same-platform, deterministic renders. For baselines that must survive a Windows-author → Linux-CI round trip, loosen both. GraphCompose's own CV / cover-letter parity tests calibrate to:
+
+```java
+PdfVisualRegression.standard()
+        .perPixelTolerance(8)            // absorb sub-pixel anti-aliasing drift
+        .mismatchedPixelBudget(50_000)   // ~glyph edges across a full A4 page
+        .assertMatchesBaseline(slug, pdfBytes);
+```
+
+Tune these to your fonts and page density: too tight and CI flakes on anti-aliasing noise; too loose and a real regression slips through. Start from the values above and tighten until CI is stable.
+
+## Using visual regression in downstream projects
+
+Library consumers use the same published helpers:
+
+```java
+import com.demcha.compose.testing.visual.PdfVisualRegression;
+
+PdfVisualRegression.standard()
+        .baselineRoot(Path.of("src", "test", "resources", "pdf-baselines"))
+        .perPixelTolerance(8)
+        .mismatchedPixelBudget(50_000)
+        .assertMatchesBaseline("reports/monthly_invoice", pdfBytes);
+```
+
+`PublicVisualApiDogfoodTest` in this repository drives exactly this consumer workflow end-to-end and proves the published surface is sufficient without any package-private access.
+
+## When not to use pixel regression
+
+- when structural geometry is what you care about → use a [layout snapshot](./layout-snapshot-testing.md) (cheaper, cross-platform stable)
+- when a small unit test proves the same rule more directly
+- as the *only* gate on a CI that runs on a different OS than where baselines were recorded — pair it with semantic snapshots and a sensible tolerance budget
+
+## Examples in this repository
+
+- `CvV2VisualParityTest`, `CoverLetterV2VisualParityTest` — preset parity with Windows-baseline / Linux-CI calibration
+- `ShapeContainerVisualRegressionTest` — engine primitive fidelity
+- `TableRowSpanDemoTest` — table rendering
+- `PdfVisualRegressionTest` — the harness's own unit tests
+- `PublicVisualApiDogfoodTest` — consumer-surface dogfood
@@ -40,7 +40,16 @@
  */
 public final class PdfVisualRegression {
 
-    private static final String APPROVE_SYS_PROP = "graphcompose.visual.approve";
+    /**
+     * System property that switches the harness into approve mode
+     * ({@code -Dgraphcompose.visual.approve=true}): instead of asserting, it
+     * (re)writes the current renders to the baseline location and skips the
+     * diff. The environment variable {@code GRAPHCOMPOSE_VISUAL_APPROVE=true}
+     * works as a fallback.
+     *
+     * @since 1.6.9
+     */
+    public static final String APPROVE_PROPERTY = "graphcompose.visual.approve";
     private static final String APPROVE_ENV_VAR = "GRAPHCOMPOSE_VISUAL_APPROVE";
 
     private final Path baselineRoot;
@@ -155,7 +164,7 @@ public void assertMatchesBaseline(String baselineName, byte[] pdfBytes) throws I
                 Path actualOut = sidecarPath(baselineName, page, "actual");
                 ImageIO.write(rendered.get(page), "png", actualOut.toFile());
                 failures.add("Missing baseline " + baseline + " — wrote rendered output to " + actualOut
-                        + ". Re-run with -D" + APPROVE_SYS_PROP + "=true to approve.");
+                        + ". Re-run with -D" + APPROVE_PROPERTY + "=true to approve.");
                 continue;
             }
             BufferedImage expected = ImageIO.read(baseline.toFile());
@@ -206,7 +215,7 @@ private Path sidecarPath(String baselineName, int pageIndex, String suffix) {
     }
 
     private static boolean approveMode() {
-        String prop = System.getProperty(APPROVE_SYS_PROP);
+        String prop = System.getProperty(APPROVE_PROPERTY);
         if (prop != null) {
             return Boolean.parseBoolean(prop);
         }

@@ -0,0 +1,88 @@
+package com.demcha.testing.visual;
+
+import com.demcha.compose.GraphCompose;
+import com.demcha.compose.document.api.DocumentPageSize;
+import com.demcha.compose.document.api.DocumentSession;
+import com.demcha.compose.document.style.DocumentTextStyle;
+import com.demcha.compose.testing.visual.PdfVisualRegression;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.nio.file.Path;
+
+import static org.assertj.core.api.Assertions.assertThat;
+
+/**
+ * Drives a consumer-style "compose a document, then pixel-test the rendered
+ * PDF" workflow entirely through the public {@code
+ * com.demcha.compose.testing.visual} surface — no package-private access.
+ *
+ * <p>Sibling to {@code LayoutSnapshotPublicApiDogfoodTest} on the semantic
+ * layer; it proves the published harness is sufficient for downstream adoption
+ * and guards against accidental private-to-public regressions in CI.</p>
+ */
+class PublicVisualApiDogfoodTest {
+
+    @TempDir
+    Path tempDir;
+
+    @Test
+    void shouldSupportConsumerPixelRegressionThroughPublicApi() throws Exception {
+        Path baselineRoot = tempDir.resolve("pdf-baselines");
+        String baselineName = "consumer_simple_document";
+        byte[] pdfBytes = renderConsumerDocument();
+
+        PdfVisualRegression harness = PdfVisualRegression.standard().baselineRoot(baselineRoot);
+
+        // 1. Consumer blesses a fresh baseline via the public approve flag.
+        String previous = System.getProperty(PdfVisualRegression.APPROVE_PROPERTY);
+        try {
+            System.setProperty(PdfVisualRegression.APPROVE_PROPERTY, "true");
+            harness.assertMatchesBaseline(baselineName, pdfBytes);
+        } finally {
+            restoreProperty(PdfVisualRegression.APPROVE_PROPERTY, previous);
+        }
+
+        assertThat(baselineRoot.resolve(baselineName + "-page-0.png"))
+                .as("approve mode should write a page-0 baseline")
+                .exists()
+                .isRegularFile();
+
+        // 2. A second render of the same document matches the just-written baseline.
+        harness.assertMatchesBaseline(baselineName, pdfBytes);
+
+        // 3. A matching run must not leave a mismatch sidecar.
+        assertThat(baselineRoot.resolve(baselineName + "-page-0.actual.png"))
+                .as("a matching run must not leave a mismatch sidecar")
+                .doesNotExist();
+    }
+
+    private byte[] renderConsumerDocument() throws Exception {
+        try (DocumentSession document = GraphCompose.document()
+                .pageSize(DocumentPageSize.A4)
+                .margin(18, 18, 18, 18)
+                .create()) {
+            document.dsl()
+                    .pageFlow()
+                    .name("ConsumerRoot")
+                    .spacing(8)
+                    .addParagraph(paragraph -> paragraph
+                            .name("Title")
+                            .text("Consumer pixel baseline")
+                            .textStyle(DocumentTextStyle.DEFAULT))
+                    .addShape(shape -> shape
+                            .name("AccentBox")
+                            .size(120, 32))
+                    .build();
+            return document.toPdfBytes();
+        }
+    }
+
+    private void restoreProperty(String key, String previous) {
+        if (previous == null) {
+            System.clearProperty(key);
+        } else {
+            System.setProperty(key, previous);
+        }
+    }
+}