Spaces:
Running
Running
| """Make the page Text-to-Scene only (user, 2026-10-01: drop Image-to-Scene from the video, the site and the figures). | |
| Edits index.html, cases data and the bar script in place. Each step asserts it found what it edits, so a re-run on an | |
| already-edited page, or on a page whose structure has moved, stops instead of silently doing half the job. | |
| python3 scripts/t2s_only.py | |
| """ | |
| from __future__ import annotations | |
| import json, re | |
| from pathlib import Path | |
| from bs4 import BeautifulSoup, NavigableString | |
| ROOT = Path(__file__).resolve().parent.parent | |
| I2S_CASES = ("comp-31-old-industrial-pallet-bay-repair-v2", "comp-61-new-york-mailbox-pair-repair-v2") | |
| def one(found, what): | |
| assert found is not None, f"not found: {what}" | |
| return found | |
| def index(): | |
| p = ROOT / "index.html" | |
| soup = BeautifulSoup(p.read_text(), "html.parser") | |
| # head | |
| one(soup.title, "title").string = "Code4Scene: Benchmarking Coding Agents for Constructing 3D Scenes" | |
| one(soup.find("meta", attrs={"name": "description"}), "meta description")["content"] = ( | |
| "A benchmark that scores the engine-native 3D scenes coding agents build in Unreal Engine from open-ended descriptions.") | |
| # intro | |
| intro = one(soup.find(id="top"), "intro") | |
| h1 = one(intro.find("h1"), "intro h1") | |
| h1.clear(); h1.append("Constructing"); h1.append(soup.new_tag("br")); em = soup.new_tag("em"); em.string = "3D Scenes with Code"; h1.append(em) | |
| one(intro.find(class_="intro-text"), "intro text").string = ( | |
| "Benchmarking coding agents that turn open-ended scene descriptions into engine-native 3D scenes in Unreal Engine. " | |
| "Measuring spatial reasoning, task fulfillment and physical validity.") | |
| cap = one(intro.find(class_="featured-video-caption"), "video caption").find("span") | |
| one(cap, "video caption span").string = "Construction and evaluation" | |
| # the teaser is v31 (no editing subtitle on its end card), under a new name so cached copies of the old cut are not reused | |
| for src in intro.find_all("source"): | |
| src["src"] = "media/teaser-v31.mp4" | |
| for a in intro.find_all("a", href="media/teaser.mp4"): | |
| a["href"] = "media/teaser-v31.mp4" | |
| # task format: the construction example alone, without a one-tab tab bar | |
| tf = one(soup.find(id="task-format"), "task format") | |
| one(tf.find(id="task-i2s"), "task-i2s").decompose() | |
| one(tf.find(class_="task-tabs"), "task tabs").decompose() | |
| # leaderboard: Text-to-Scene only (the overall score is half Image-to-Scene) | |
| lb = one(soup.find(id="leaderboard"), "leaderboard") | |
| one(lb.find(class_="lb-tabs"), "leaderboard tabs").decompose() | |
| for el in lb.find_all(attrs={"data-view": True}): | |
| if el.get("data-view") in ("overall", "i2s"): | |
| el.decompose() | |
| else: | |
| st = (el.get("style") or "").replace("display:none", "").strip(" ;") | |
| if st: el["style"] = st | |
| else: del el["style"] | |
| desc = one(lb.find(class_="section-desc"), "leaderboard desc") | |
| desc.string = "14 coding-agent configurations · paper results on the original 20 public construction cases." | |
| sub = one(lb.find(class_="figure-subtitle"), "pareto subtitle") | |
| sub.string = "Quality against the cost of one case · original 20-case public evaluation" | |
| for t in lb.find_all(class_="figure-caption"): | |
| t.string = re.sub(r"\s*Cost is the mean USD per case; ov.*$", " Cost is the mean USD per case.", t.get_text()) | |
| # cases: the four construction cases | |
| cs = one(soup.find(id="cases"), "cases") | |
| removed = 0 | |
| for scene in cs.find_all(attrs={"data-case": True}): | |
| if scene["data-case"] in I2S_CASES: | |
| card = scene | |
| while card.parent is not None and card.parent.name != "body" and card.parent.get("id") != "cases" and \ | |
| not any(c for c in (card.parent.get("class") or []) if "grid" in c): | |
| card = card.parent | |
| card.decompose(); removed += 1 | |
| assert removed == 2, removed | |
| for p_ in cs.find_all("p"): | |
| if "Compare repairs" in p_.get_text(): | |
| p_.string = "Drag any scene to inspect the saved 3D output, or open a case to explore every agent." | |
| # takeaways: Spatial Composition only | |
| tk = one(soup.find(id="takeaways"), "takeaways") | |
| kept = 0 | |
| for card in tk.find_all(class_="tk"): | |
| h = " ".join(one(card.find("h3"), "takeaway h3").get_text(" ").split()) | |
| if h.startswith("Spatial Composition remains"): | |
| kept += 1 | |
| for n in card.find_all(class_="n"): | |
| n.string = "Takeaway" | |
| else: | |
| card.decompose() | |
| assert kept == 1, kept | |
| for h in tk.find_all("h2"): | |
| if h.get_text(strip=True) == "Takeaways": | |
| h.string = "Takeaway" | |
| # the case-count strip: construction counts only | |
| av = one(soup.find(class_="availability"), "availability strip") | |
| spans = av.find_all("span", recursive=False); assert len(spans) == 2 | |
| spans[0].b.string = "160"; spans[0].b.next_sibling.replace_with(" construction cases in the full benchmark target, including private cases") | |
| spans[1].b.string = "129"; spans[1].b.next_sibling.replace_with(" public construction cases") | |
| # nav label | |
| for a in soup.find_all("a"): | |
| if a.get_text(strip=True) == "Takeaways": | |
| a.string = "Takeaway" | |
| p.write_text(str(soup)) | |
| left = re.findall(r"(?i)image-to-scene|\bi2s\b|repair|corrupted", p.read_text()) | |
| print("index.html: remaining editing mentions:", len(left)) | |
| def bars(): | |
| p = ROOT / "assets/leaderboard-bars.js" | |
| s = p.read_text() | |
| m = re.search(r"const ALL = (\{.*?\});\n", s, re.S) | |
| data = json.loads(one(m, "const ALL").group(1)) | |
| assert set(data) >= {"t2s"}, data.keys() | |
| s = s[:m.start(1)] + json.dumps({"t2s": data["t2s"]}) + s[m.end(1):] | |
| p.write_text(s) | |
| print("leaderboard-bars.js: views", list(json.loads(re.search(r"const ALL = (\{.*?\});\n", s, re.S).group(1)))) | |
| def site_data(): | |
| p = ROOT / "data/site.js" | |
| S = json.loads(p.read_text().split("=", 1)[1].rstrip().rstrip(";")) | |
| before = len(S["cases"]) | |
| S["cases"] = [c for c in S["cases"] if c["track"] == "t2s"] | |
| p.write_text("window.SITE = " + json.dumps(S) + ";\n") | |
| print("data/site.js: cases", before, "->", len(S["cases"])) | |
| # the cases page states the counts in its header | |
| n3d = sum(1 for c in S["cases"] for a in c["agents"] if a.get("glb")) + sum(1 for c in S["cases"] for k in ("gt_glb", "input_glb") if c.get(k)) | |
| cp = ROOT / "cases.html"; cs = cp.read_text() | |
| cs = re.sub(r"\d+ cases, \d+ scenes in 3D", f'{len(S["cases"])} cases, {n3d} scenes in 3D', cs); cp.write_text(cs) | |
| # the packed model files of the two Image-to-Scene cases (repairs, ground truth, corrupted input), from data/site.js before this edit | |
| I2S_MODELS = ["38071c69a86d174dc55f.bin", "46734878f740aae816d8.bin", "486b4742affe835723a3.bin", "66da6868d20ebb100197.bin", "6afbbaec6b7a4271b21f.bin", "7daf8107bb630cffdc18.bin", "c79715b9e6967c4f6c90.bin", "db5c0c5d2b78f9e47bb7.bin", "e6a6333939ccfe57d584.bin", "e9177d0cbf727df88100.bin", "f2b2b285e27a26ee6780.bin", "f3648a5ad9887ac0937d.bin"] | |
| def progressive(): | |
| p = ROOT / "data/progressive-scenes.json" | |
| d = json.loads(p.read_text()); n0 = len(d) | |
| d = {k: v for k, v in d.items() if k not in I2S_MODELS} | |
| p.write_text(json.dumps(d, indent=2) + "\n") | |
| print("data/progressive-scenes.json:", n0, "->", len(d)) | |
| if __name__ == "__main__": | |
| index(); bars(); site_data(); progressive() | |