diff --git a/README.md b/README.md
index 7aac044..20cff8b 100644
--- a/README.md
+++ b/README.md
@@ -10,7 +10,7 @@ Instead of pretending this is a resize problem, Repixelizer treats it as lattice
 
 Repixelizer was built to rescue fake pixel art, but it can also be used to generate pixel art directly from non-pixel source art when the shapes are clean and the local structure is doing something useful.
 
-Top row: glossy non-pixel badge art repixelized at `128x128`. Bottom row: an AI fake-pixel interpretation of the same badge cleaned up with automatic lattice inference, which landed on `122x122` for this image.
+Rows show the source art, a plain Lanczos downscale to the same target size, and the repixelized result. The bottom row also carries a sword-guard picture-in-picture, because that tiny region is one of the more honest tests of whether the tool is preserving local cell adjacency instead of just getting the vibes approximately correct.
 
 ![Repixelizer example comparison](docs/readme-assets/badge-example-sheet.png)
 
@@ -118,6 +118,16 @@ Run the focused test suite with:
 .venv\Scripts\python -m pytest -q
 ```
 
+## Regenerating README Assets
+
+The README comparison sheet is generated from repo-tracked fixtures, not from random artifacts left lying around:
+
+```powershell
+.venv\Scripts\python scripts\generate_readme_previews.py --vector-input tests\fixtures\real\ai-badge-vector.png --ai-input tests\fixtures\real\ai-badge-cleaned.png --out-sheet docs\readme-assets\badge-example-sheet.png --out-guard-crop docs\readme-assets\guard-right-crop-comparison.png --scratch-dir artifacts\readme-build --device cpu
+```
+
+That regenerates the README sheet, the standalone sword-guard closeup strip, and scratch outputs under `artifacts/readme-build/` so you can inspect the actual low-res results used to build the docs sample.
+
 ## Diagnostic Closeups
 
 For docs or regression notes, you can generate reproducible source/output closeups from output-grid cell coordinates with:
diff --git a/docs/readme-assets/badge-example-sheet.png b/docs/readme-assets/badge-example-sheet.png
index 1ecaacd..55573c2 100644
Binary files a/docs/readme-assets/badge-example-sheet.png and b/docs/readme-assets/badge-example-sheet.png differ
diff --git a/docs/readme-assets/guard-right-crop-comparison.png b/docs/readme-assets/guard-right-crop-comparison.png
new file mode 100644
index 0000000..ce8491a
Binary files /dev/null and b/docs/readme-assets/guard-right-crop-comparison.png differ
diff --git a/scripts/generate_readme_previews.py b/scripts/generate_readme_previews.py
new file mode 100644
index 0000000..bfa2d7c
--- /dev/null
+++ b/scripts/generate_readme_previews.py
@@ -0,0 +1,386 @@
+from __future__ import annotations
+
+import argparse
+import json
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+from PIL import Image, ImageDraw
+
+from repixelizer.baselines import lanczos_resize_baseline
+from repixelizer.io import load_rgba, nearest_resize, save_rgba
+from repixelizer.pipeline import run_pipeline
+
+
+@dataclass(slots=True)
+class RenderedPanel:
+    image: Image.Image
+    content_box: tuple[int, int, int, int]
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Regenerate the README comparison sheet and guard closeup from repo fixtures.")
+    parser.add_argument(
+        "--vector-input",
+        default="tests/fixtures/real/ai-badge-vector.png",
+        help="High-resolution non-pixel badge source.",
+    )
+    parser.add_argument(
+        "--ai-input",
+        default="tests/fixtures/real/ai-badge-cleaned.png",
+        help="Cleaned AI fake-pixel badge source.",
+    )
+    parser.add_argument(
+        "--out-sheet",
+        default="docs/readme-assets/badge-example-sheet.png",
+        help="Output path for the full README comparison sheet.",
+    )
+    parser.add_argument(
+        "--out-guard-crop",
+        default="docs/readme-assets/guard-right-crop-comparison.png",
+        help="Output path for the standalone sword-guard closeup strip.",
+    )
+    parser.add_argument(
+        "--scratch-dir",
+        default="artifacts/readme-build",
+        help="Directory for intermediate low-res outputs and diagnostics.",
+    )
+    parser.add_argument(
+        "--vector-target-size",
+        type=int,
+        default=128,
+        help="Target size for the non-pixel source repixelization run.",
+    )
+    parser.add_argument(
+        "--guard-cell-bbox",
+        nargs=4,
+        type=int,
+        default=(73, 20, 107, 44),
+        metavar=("X0", "Y0", "X1", "Y1"),
+        help="Sword-guard crop bounds in output-grid cell coordinates for the AI row.",
+    )
+    parser.add_argument("--steps", type=int, default=48, help="Optimizer step budget for both runs.")
+    parser.add_argument("--device", default="cpu", choices=("auto", "cpu", "cuda"), help="Torch device for generation.")
+    return parser.parse_args()
+
+
+def _to_image(rgba: np.ndarray) -> Image.Image:
+    rgba8 = np.clip(np.rint(rgba * 255.0), 0, 255).astype(np.uint8)
+    return Image.fromarray(rgba8, mode="RGBA")
+
+
+def _checkerboard(width: int, height: int, tile: int = 12) -> Image.Image:
+    image = Image.new("RGBA", (width, height), (236, 236, 236, 255))
+    draw = ImageDraw.Draw(image)
+    accent = (218, 218, 218, 255)
+    for y in range(0, height, tile):
+        for x in range(0, width, tile):
+            if ((x // tile) + (y // tile)) % 2:
+                draw.rectangle((x, y, min(width, x + tile), min(height, y + tile)), fill=accent)
+    return image
+
+
+def _fit_rgba(
+    rgba: np.ndarray,
+    *,
+    target_width: int,
+    target_height: int,
+    pixelated: bool,
+) -> tuple[Image.Image, tuple[int, int, int, int]]:
+    image = _to_image(rgba)
+    scale = min(target_width / max(1, image.width), target_height / max(1, image.height))
+    resized_width = max(1, int(round(image.width * scale)))
+    resized_height = max(1, int(round(image.height * scale)))
+    resample = Image.Resampling.NEAREST if pixelated else Image.Resampling.LANCZOS
+    resized = image.resize((resized_width, resized_height), resample=resample)
+    offset_x = (target_width - resized_width) // 2
+    offset_y = (target_height - resized_height) // 2
+    return resized, (offset_x, offset_y, resized_width, resized_height)
+
+
+def _render_panel(
+    label: str,
+    *,
+    rgba: np.ndarray,
+    pixelated: bool,
+    panel_width: int,
+    panel_height: int,
+    subtitle: str | None = None,
+) -> RenderedPanel:
+    header_height = 38 if subtitle else 28
+    gutter = 10
+    canvas = Image.new("RGBA", (panel_width, panel_height), (18, 18, 20, 255))
+    checker = _checkerboard(panel_width - gutter * 2, panel_height - header_height - gutter * 2)
+    canvas.alpha_composite(checker, (gutter, header_height + gutter))
+
+    fitted, (inner_x, inner_y, inner_w, inner_h) = _fit_rgba(
+        rgba,
+        target_width=checker.width,
+        target_height=checker.height,
+        pixelated=pixelated,
+    )
+    content_x = gutter + inner_x
+    content_y = header_height + gutter + inner_y
+    canvas.alpha_composite(fitted, (content_x, content_y))
+
+    draw = ImageDraw.Draw(canvas)
+    draw.text((gutter, 6), label, fill=(255, 255, 255, 255))
+    if subtitle:
+        draw.text((gutter, 22), subtitle, fill=(200, 200, 200, 255))
+    draw.rounded_rectangle((0, 0, panel_width - 1, panel_height - 1), radius=10, outline=(72, 72, 80, 255), width=1)
+    return RenderedPanel(canvas, (content_x, content_y, inner_w, inner_h))
+
+
+def _draw_source_bbox(render: RenderedPanel, source_bbox: tuple[int, int, int, int], *, source_width: int, source_height: int) -> None:
+    x0, y0, x1, y1 = source_bbox
+    inner_x, inner_y, inner_w, inner_h = render.content_box
+    left = inner_x + int(round(inner_w * (x0 / max(1, source_width))))
+    right = inner_x + int(round(inner_w * (x1 / max(1, source_width))))
+    top = inner_y + int(round(inner_h * (y0 / max(1, source_height))))
+    bottom = inner_y + int(round(inner_h * (y1 / max(1, source_height))))
+    draw = ImageDraw.Draw(render.image)
+    for delta, color in ((0, (255, 245, 120, 255)), (1, (20, 20, 20, 255))):
+        draw.rectangle((left - delta, top - delta, right + delta, bottom + delta), outline=color, width=1)
+
+
+def _build_guard_strip(
+    source_crop: np.ndarray,
+    lanczos_crop: np.ndarray,
+    repixelized_crop: np.ndarray,
+    *,
+    include_lanczos: bool = True,
+    panel_width: int = 220,
+    panel_height: int = 170,
+    title: str = "Sword guard close-up",
+) -> Image.Image:
+    panels = [
+        _render_panel("Source", rgba=source_crop, pixelated=True, panel_width=panel_width, panel_height=panel_height),
+        _render_panel("Repixelized", rgba=repixelized_crop, pixelated=True, panel_width=panel_width, panel_height=panel_height),
+    ]
+    if include_lanczos:
+        panels.insert(
+            1,
+            _render_panel("Lanczos", rgba=lanczos_crop, pixelated=True, panel_width=panel_width, panel_height=panel_height),
+        )
+    title_height = 24
+    gap = 10
+    width = sum(panel.image.width for panel in panels) + gap * (len(panels) + 1)
+    height = title_height + max(panel.image.height for panel in panels) + gap * 2
+    canvas = Image.new("RGBA", (width, height), (14, 14, 16, 242))
+    draw = ImageDraw.Draw(canvas)
+    draw.rounded_rectangle((0, 0, width - 1, height - 1), radius=12, fill=(14, 14, 16, 242), outline=(92, 92, 104, 255), width=1)
+    draw.text((gap, 4), title, fill=(255, 255, 255, 255))
+    x = gap
+    for panel in panels:
+        canvas.alpha_composite(panel.image, (x, title_height + gap))
+        x += panel.image.width + gap
+    return canvas
+
+
+def _cell_bbox_to_source_bbox(
+    cell_bbox: tuple[int, int, int, int],
+    *,
+    source_width: int,
+    source_height: int,
+    target_width: int,
+    target_height: int,
+) -> tuple[int, int, int, int]:
+    x0, y0, x1, y1 = cell_bbox
+    return (
+        int(np.floor(x0 * source_width / max(1, target_width))),
+        int(np.floor(y0 * source_height / max(1, target_height))),
+        int(np.ceil(x1 * source_width / max(1, target_width))),
+        int(np.ceil(y1 * source_height / max(1, target_height))),
+    )
+
+
+def _crop_rgba(rgba: np.ndarray, bbox: tuple[int, int, int, int]) -> np.ndarray:
+    x0, y0, x1, y1 = bbox
+    return rgba[y0:y1, x0:x1]
+
+
+def _save_summary(path: Path, payload: dict[str, object]) -> None:
+    path.write_text(json.dumps(payload, indent=2), encoding="utf-8")
+
+
+def main() -> None:
+    args = _parse_args()
+    scratch_dir = Path(args.scratch_dir)
+    scratch_dir.mkdir(parents=True, exist_ok=True)
+
+    vector_input = Path(args.vector_input)
+    ai_input = Path(args.ai_input)
+    out_sheet = Path(args.out_sheet)
+    out_guard_crop = Path(args.out_guard_crop)
+    out_sheet.parent.mkdir(parents=True, exist_ok=True)
+    out_guard_crop.parent.mkdir(parents=True, exist_ok=True)
+
+    vector_result = run_pipeline(
+        vector_input,
+        scratch_dir / "badge-vector-repixelized.png",
+        target_size=args.vector_target_size,
+        diagnostics_dir=scratch_dir / "vector-diag",
+        steps=args.steps,
+        device=args.device,
+    )
+    ai_result = run_pipeline(
+        ai_input,
+        scratch_dir / "badge-ai-repixelized.png",
+        diagnostics_dir=scratch_dir / "ai-diag",
+        steps=args.steps,
+        device=args.device,
+    )
+
+    vector_source = load_rgba(vector_input)
+    ai_source = load_rgba(ai_input)
+    vector_lanczos = lanczos_resize_baseline(vector_source, width=args.vector_target_size, height=args.vector_target_size)
+    ai_lanczos = lanczos_resize_baseline(
+        ai_source,
+        width=ai_result.inference.target_width,
+        height=ai_result.inference.target_height,
+    )
+
+    save_rgba(scratch_dir / "badge-vector-lanczos.png", vector_lanczos)
+    save_rgba(scratch_dir / "badge-ai-lanczos.png", ai_lanczos)
+
+    panel_width = 320
+    panel_height = 360
+    row_gap = 18
+    col_gap = 18
+    margin = 24
+    row_title_height = 22
+
+    vector_panels = [
+        _render_panel(
+            "Source",
+            rgba=vector_source,
+            pixelated=False,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle="Glossy badge art",
+        ),
+        _render_panel(
+            "Lanczos",
+            rgba=vector_lanczos,
+            pixelated=True,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle=f"{args.vector_target_size}x{args.vector_target_size}",
+        ),
+        _render_panel(
+            "Repixelized",
+            rgba=vector_result.output_rgba,
+            pixelated=True,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle=f"{vector_result.inference.target_width}x{vector_result.inference.target_height}",
+        ),
+    ]
+    ai_panels = [
+        _render_panel(
+            "Source",
+            rgba=ai_source,
+            pixelated=False,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle="Cleaned AI fake pixel art",
+        ),
+        _render_panel(
+            "Lanczos",
+            rgba=ai_lanczos,
+            pixelated=True,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle=f"{ai_result.inference.target_width}x{ai_result.inference.target_height}",
+        ),
+        _render_panel(
+            "Repixelized",
+            rgba=ai_result.output_rgba,
+            pixelated=True,
+            panel_width=panel_width,
+            panel_height=panel_height,
+            subtitle=f"{ai_result.inference.target_width}x{ai_result.inference.target_height} auto",
+        ),
+    ]
+
+    guard_bbox = _cell_bbox_to_source_bbox(
+        tuple(args.guard_cell_bbox),
+        source_width=ai_source.shape[1],
+        source_height=ai_source.shape[0],
+        target_width=ai_result.inference.target_width,
+        target_height=ai_result.inference.target_height,
+    )
+    ai_lanczos_preview = nearest_resize(ai_lanczos, width=ai_source.shape[1], height=ai_source.shape[0])
+    ai_repixelized_preview = nearest_resize(ai_result.output_rgba, width=ai_source.shape[1], height=ai_source.shape[0])
+    for panel in ai_panels:
+        _draw_source_bbox(
+            panel,
+            guard_bbox,
+            source_width=ai_source.shape[1],
+            source_height=ai_source.shape[0],
+        )
+
+    guard_strip = _build_guard_strip(
+        _crop_rgba(ai_source, guard_bbox),
+        _crop_rgba(ai_lanczos_preview, guard_bbox),
+        _crop_rgba(ai_repixelized_preview, guard_bbox),
+    )
+    guard_strip.save(out_guard_crop)
+    guard_inset = _build_guard_strip(
+        _crop_rgba(ai_source, guard_bbox),
+        _crop_rgba(ai_lanczos_preview, guard_bbox),
+        _crop_rgba(ai_repixelized_preview, guard_bbox),
+        include_lanczos=False,
+        panel_width=190,
+        panel_height=142,
+        title="Sword guard",
+    )
+
+    width = margin * 2 + panel_width * 3 + col_gap * 2
+    height = margin * 2 + row_title_height * 2 + panel_height * 2 + row_gap
+    canvas = Image.new("RGBA", (width, height), (11, 11, 13, 255))
+    draw = ImageDraw.Draw(canvas)
+    draw.text((margin, margin - 2), "Non-pixel source -> 128x128", fill=(255, 255, 255, 255))
+    draw.text((margin, margin + row_title_height + panel_height + row_gap - 2), "AI fake pixel art -> auto lattice", fill=(255, 255, 255, 255))
+
+    row1_y = margin + row_title_height
+    row2_y = row1_y + panel_height + row_gap + row_title_height
+    x = margin
+    for panel in vector_panels:
+        canvas.alpha_composite(panel.image, (x, row1_y))
+        x += panel_width + col_gap
+    x = margin
+    for panel in ai_panels:
+        canvas.alpha_composite(panel.image, (x, row2_y))
+        x += panel_width + col_gap
+
+    inset_x = width - guard_inset.width - margin
+    inset_y = row2_y + panel_height - guard_inset.height - 18
+    canvas.alpha_composite(guard_inset, (inset_x, inset_y))
+
+    out_sheet.parent.mkdir(parents=True, exist_ok=True)
+    canvas.save(out_sheet)
+
+    _save_summary(
+        scratch_dir / "readme-preview-summary.json",
+        {
+            "vector_input": str(vector_input),
+            "ai_input": str(ai_input),
+            "vector_target_size": args.vector_target_size,
+            "ai_inferred_width": ai_result.inference.target_width,
+            "ai_inferred_height": ai_result.inference.target_height,
+            "guard_cell_bbox": list(args.guard_cell_bbox),
+            "guard_source_bbox": list(guard_bbox),
+            "out_sheet": str(out_sheet),
+            "out_guard_crop": str(out_guard_crop),
+            "scratch_dir": str(scratch_dir),
+        },
+    )
+
+    print(str(out_sheet))
+    print(str(out_guard_crop))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/repixelizer/baselines.py b/src/repixelizer/baselines.py
index a26f67a..85bbc74 100644
--- a/src/repixelizer/baselines.py
+++ b/src/repixelizer/baselines.py
@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import numpy as np
 
-from .io import box_resize
+from .io import box_resize, lanczos_resize
 from .palette import derive_palette
 
 
@@ -10,6 +10,10 @@ def naive_resize_baseline(rgba: np.ndarray, width: int, height: int) -> np.ndarr
     return box_resize(rgba, width=width, height=height)
 
 
+def lanczos_resize_baseline(rgba: np.ndarray, width: int, height: int) -> np.ndarray:
+    return lanczos_resize(rgba, width=width, height=height)
+
+
 def _nearest_palette_color(pixel: np.ndarray, palette: np.ndarray) -> np.ndarray:
     distances = np.sum((palette - pixel[None, :]) ** 2, axis=1)
     return palette[int(np.argmin(distances))]
diff --git a/src/repixelizer/compare.py b/src/repixelizer/compare.py
index d9bcc34..8d61221 100644
--- a/src/repixelizer/compare.py
+++ b/src/repixelizer/compare.py
@@ -6,7 +6,7 @@ from typing import Any
 import numpy as np
 from PIL import Image, ImageDraw
 
-from .baselines import error_diffusion_baseline, naive_resize_baseline
+from .baselines import error_diffusion_baseline, lanczos_resize_baseline, naive_resize_baseline
 from .diagnostics import write_compare_csv, write_json
 from .io import nearest_resize, save_rgba
 from .metrics import (
@@ -30,6 +30,7 @@ def _to_uint8(rgba: np.ndarray) -> np.ndarray:
 def _contact_sheet(
     source: np.ndarray,
     optimized: np.ndarray,
+    lanczos: np.ndarray,
     naive: np.ndarray,
     diffusion: np.ndarray,
     path: str | Path,
@@ -38,6 +39,7 @@ def _contact_sheet(
     panels = [
         ("Source", source),
         ("Optimized", nearest_resize(optimized, width=width, height=height)),
+        ("Lanczos", nearest_resize(lanczos, width=width, height=height)),
         ("Naive", nearest_resize(naive, width=width, height=height)),
         ("Diffusion", nearest_resize(diffusion, width=width, height=height)),
     ]
@@ -79,6 +81,7 @@ def run_compare(
     )
     source = result.source_rgba
     palette = load_palette(palette_path) if palette_path else None
+    lanczos = lanczos_resize_baseline(source, width=result.inference.target_width, height=result.inference.target_height)
     naive = naive_resize_baseline(source, width=result.inference.target_width, height=result.inference.target_height)
     diffusion = error_diffusion_baseline(
         source,
@@ -86,10 +89,11 @@ def run_compare(
         height=result.inference.target_height,
         palette=palette,
     )
-    _contact_sheet(source, result.output_rgba, naive, diffusion, diagnostics_path / "compare-sheet.png")
+    _contact_sheet(source, result.output_rgba, lanczos, naive, diffusion, diagnostics_path / "compare-sheet.png")
     rows = []
     for name, image in (
         ("optimized", result.output_rgba),
+        ("lanczos", lanczos),
         ("naive", naive),
         ("diffusion", diffusion),
     ):
diff --git a/src/repixelizer/io.py b/src/repixelizer/io.py
index 19752d8..c0a2401 100644
--- a/src/repixelizer/io.py
+++ b/src/repixelizer/io.py
@@ -45,3 +45,9 @@ def bilinear_resize(rgba: np.ndarray, width: int, height: int) -> np.ndarray:
     image = Image.fromarray(np.clip(np.rint(rgba * 255.0), 0, 255).astype(np.uint8), mode="RGBA")
     resized = image.resize((width, height), resample=Image.Resampling.BILINEAR)
     return np.asarray(resized, dtype=np.float32) / 255.0
+
+
+def lanczos_resize(rgba: np.ndarray, width: int, height: int) -> np.ndarray:
+    image = Image.fromarray(np.clip(np.rint(rgba * 255.0), 0, 255).astype(np.uint8), mode="RGBA")
+    resized = image.resize((width, height), resample=Image.Resampling.LANCZOS)
+    return np.asarray(resized, dtype=np.float32) / 255.0
diff --git a/tests/fixtures/real/ai-badge-vector.png b/tests/fixtures/real/ai-badge-vector.png
new file mode 100644
index 0000000..06d23d0
Binary files /dev/null and b/tests/fixtures/real/ai-badge-vector.png differ
diff --git a/tests/test_compare.py b/tests/test_compare.py
index a32414f..6d4d4b6 100644
--- a/tests/test_compare.py
+++ b/tests/test_compare.py
@@ -19,4 +19,4 @@ def test_compare_writes_comparison_artifacts(tmp_path: Path) -> None:
     assert output_path.exists()
     assert (diagnostics_dir / "compare-sheet.png").exists()
     assert (diagnostics_dir / "compare.csv").exists()
-    assert len(payload["rows"]) == 3
+    assert len(payload["rows"]) == 4