""" Case Study: Generate side-by-side visual comparisons. Selects representative examples and creates HTML comparison pages. Usage: python scripts/step_case_study.py """ import json import os import sys from pathlib import Path import numpy as np from PIL import Image PROJECT_ROOT = Path(__file__).parent.parent sys.path.insert(0, str(PROJECT_ROOT)) METHODS_TO_COMPARE = [ "deepseek_tiny", "deepseek_base", "deepseek_large", "qwen3_256", "qwen3_1k", "qwen3_full", ] METHOD_LABELS = { "deepseek_tiny": "DeepSeek-OCR tiny (73 tok)", "deepseek_base": "DeepSeek-OCR base (273 tok)", "deepseek_large": "DeepSeek-OCR large (421 tok)", "qwen3_256": "Qwen3-VL 256 (722 tok)", "qwen3_1k": "Qwen3-VL 1k (3043 tok)", "qwen3_full": "Qwen3-VL full (6746 tok)", } def select_representative_samples(benchmark_dir, n=8): """Select diverse, representative samples based on CLIP score variance.""" clips_by_sample = {} for method in METHODS_TO_COMPARE: clip_file = Path(benchmark_dir) / method / "clip_scores.json" if not clip_file.exists(): continue with open(clip_file) as f: data = json.load(f) per_sample = data.get("per_sample", {}) for sid, val in per_sample.items(): score = val.get("clip_score", val) if isinstance(val, dict) else float(val) clips_by_sample.setdefault(sid, {})[method] = score candidates = [] for sid, scores in clips_by_sample.items(): if len(scores) < 4: continue vals = list(scores.values()) candidates.append({ "id": sid, "mean_clip": np.mean(vals), "std_clip": np.std(vals), "max_clip": max(vals), "min_clip": min(vals), "range": max(vals) - min(vals), "scores": scores, }) candidates.sort(key=lambda c: -c["range"]) selected = [] high_quality = [c for c in candidates if c["mean_clip"] > 0.85] if high_quality: selected.append(high_quality[0]) low_quality = [c for c in candidates if c["mean_clip"] < 0.65 and c not in selected] if low_quality: selected.append(low_quality[0]) high_variance = [c for c in candidates if c not in selected] high_variance.sort(key=lambda c: -c["range"]) for c in high_variance[:3]: if c not in selected: selected.append(c) mid_range = [c for c in candidates if 0.70 < c["mean_clip"] < 0.80 and c not in selected] mid_range.sort(key=lambda c: -c["range"]) for c in mid_range[:3]: if c not in selected: selected.append(c) return selected[:n] def render_html_to_png(html_path, output_path, width=1280, height=1024): """Render HTML to PNG screenshot.""" try: from playwright.sync_api import sync_playwright with sync_playwright() as p: browser = p.chromium.launch(headless=True, args=['--no-sandbox', '--disable-gpu']) page = browser.new_page(viewport={"width": width, "height": height}) page.goto(f"file://{html_path}", wait_until="networkidle", timeout=15000) page.wait_for_timeout(500) page.screenshot(path=str(output_path), full_page=False) browser.close() return True except Exception as e: print(f" Render failed: {e}") return False def generate_case_study_html(selected, benchmark_dir, ref_dir, output_dir): """Generate an HTML page with side-by-side comparisons.""" output_dir = Path(output_dir) output_dir.mkdir(parents=True, exist_ok=True) images_dir = output_dir / "images" images_dir.mkdir(exist_ok=True) for sample in selected: sid = sample["id"] ref_src = Path(ref_dir) / f"{sid}.png" if ref_src.exists(): ref_dst = images_dir / f"ref_{sid}.png" if not ref_dst.exists(): img = Image.open(ref_src) img.thumbnail((640, 800)) img.save(str(ref_dst)) for method in METHODS_TO_COMPARE: html_path = Path(benchmark_dir) / method / "html_predictions" / f"{sid}.html" render_path = images_dir / f"{method}_{sid}.png" if html_path.exists() and not render_path.exists(): print(f" Rendering {method}/{sid}...") ok = render_html_to_png(str(html_path.resolve()), str(render_path)) if ok: img = Image.open(render_path) img.thumbnail((640, 800)) img.save(str(render_path)) rows_html = [] for i, sample in enumerate(selected): sid = sample["id"] scores_str = " | ".join( f"{METHOD_LABELS.get(m, m).split('(')[0].strip()}: {sample['scores'].get(m, 'N/A'):.3f}" if isinstance(sample['scores'].get(m), float) else f"{m}: N/A" for m in METHODS_TO_COMPARE ) cells = [f'ref
Original'] for method in METHODS_TO_COMPARE: label = METHOD_LABELS.get(method, method) clip = sample["scores"].get(method) clip_str = f"CLIP: {clip:.3f}" if clip else "N/A" img_file = f"images/{method}_{sid}.png" cells.append(f'{method}
{label}
{clip_str}') row = f""" Case {i+1} (Sample ID: {sid}) — Mean CLIP: {sample['mean_clip']:.3f}, Range: {sample['range']:.3f} {''.join(cells)} """ rows_html.append(row) html = f""" UIPress Case Study

UIPress: Visual Token Compression Case Study

Side-by-side comparison of {len(selected)} representative examples across {len(METHODS_TO_COMPARE)} methods.

{''.join(rows_html)}
""" output_file = output_dir / "case_study.html" output_file.write_text(html) print(f"Case study saved to {output_file}") summary = { "n_cases": len(selected), "methods": METHODS_TO_COMPARE, "cases": [{ "id": s["id"], "mean_clip": round(s["mean_clip"], 4), "clip_range": round(s["range"], 4), "scores": {k: round(v, 4) for k, v in s["scores"].items()}, } for s in selected], } with open(output_dir / "case_study_summary.json", "w") as f: json.dump(summary, f, indent=2) return output_file def main(): benchmark_dir = PROJECT_ROOT / "results" / "benchmark" ref_dir = PROJECT_ROOT / "data" / "ref_screenshots" output_dir = PROJECT_ROOT / "results" / "case_study" print("Selecting representative samples...") selected = select_representative_samples(str(benchmark_dir), n=8) print(f"\nSelected {len(selected)} cases:") for s in selected: print(f" ID={s['id']}: mean_clip={s['mean_clip']:.3f}, range={s['range']:.3f}") print("\nGenerating case study...") output_file = generate_case_study_html(selected, str(benchmark_dir), str(ref_dir), str(output_dir)) print(f"\nDone! Open {output_file} in a browser to view.") if __name__ == "__main__": main()