update

.gitignore

@@ -1,5 +1,6 @@
 input/
 output/
+runs
 Paper2Poster/assets/
 Paper2Video/assets/
 posterbuilder/latex_proj/figures/

app.py

@@ -1,76 +1,312 @@
 import gradio as gr
-import subprocess, shutil, os, zipfile, datetime, sys, time
+import subprocess, shutil, os, zipfile, datetime, sys, time, uuid, stat, re
 from pathlib import Path
 
+# =====================
+# Version guard
+# =====================
 def _ensure_versions():
     import importlib, subprocess, sys
-    def ver(pkg):
+
+    def get_version(pkg):
         try:
             m = importlib.import_module(pkg)
             return getattr(m, "__version__", "0")
         except Exception:
             return "0"
 
-    hub_ok = False
     try:
         from packaging.version import Version
-        hv = Version(ver("huggingface_hub"))
-        hub_ok = Version("0.24.0") <= hv < Version("1.0.0")
-    except Exception:
-        pass
+    except ImportError:
+        # 安装packaging，确保下面版本比较能用
+        subprocess.check_call([sys.executable, "-m", "pip", "install", "packaging"])
+        from packaging.version import Version
+
+    # 检查 huggingface_hub
+    hub_ver = get_version("huggingface_hub")
+    hv = Version(hub_ver)
+
+    required_min = Version("0.24.0")
+    required_max = Version("1.0.0")
+
+    hub_ok = required_min <= hv < required_max
 
     if not hub_ok:
-        subprocess.check_call([sys.executable, "-m", "pip",
-                               "install", "huggingface-hub==0.27.1",
-                               "transformers==4.48.0",
-                               "--force-reinstall", "--no-deps"])
+        print(f"[INFO] huggingface_hub=={hub_ver} not in range "
+              f"[{required_min}, {required_max}), reinstalling...")
+        subprocess.check_call([
+            sys.executable, "-m", "pip", "install",
+            "huggingface-hub==0.27.1",
+            "transformers==4.48.0",
+            "--force-reinstall", "--no-deps"
+        ])
+    else:
+        print(f"[INFO] huggingface_hub version OK: {hub_ver}")
+
 _ensure_versions()
 
+# =====================
+# Paths (read-only repo root; DO NOT write here)
+# =====================
 ROOT = Path(__file__).resolve().parent
-OUTPUT_DIR = ROOT / "output"
-INPUT_DIR = ROOT / "input"
-LOGO_DIR = INPUT_DIR / "logo"
-POSTER_LATEX_DIR = ROOT / "posterbuilder" / "latex_proj"
-ZIP_PATH = ROOT / "output.zip"
-LOG_PATH = ROOT / "last_run.log"
+RUNS_DIR = ROOT / "runs"              # all per-run workspaces live here
+RUNS_DIR.mkdir(parents=True, exist_ok=True)
+
 TIMEOUT_SECONDS = 1800  # 30 minutes
+RETENTION_HOURS = 12    # auto-clean runs older than N hours
 
-def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
-    start_time = datetime.datetime.now()
-    logs = [f"🚀 Starting pipeline at {start_time.strftime('%Y-%m-%d %H:%M:%S')}\n"]
+# ---------------------
+# Utils
+# ---------------------
+def _now_str():
+    return datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
 
-    # ====== Prepare directories ======
-    for d in [OUTPUT_DIR, LOGO_DIR, POSTER_LATEX_DIR, INPUT_DIR]:
-        d.mkdir(parents=True, exist_ok=True)
+def _write_logs(log_path: Path, logs):
+    try:
+        log_path.parent.mkdir(parents=True, exist_ok=True)
+        with open(log_path, "w", encoding="utf-8") as f:
+            f.write("\n".join(logs))
+    except Exception:
+        pass
+
+def _on_rm_error(func, path, exc_info):
+    # fix "PermissionError: [Errno 13] Permission denied" for readonly files
+    os.chmod(path, stat.S_IWRITE)
+    func(path)
+
+def _copytree(src: Path, dst: Path, symlinks=True, ignore=None):
+    if dst.exists():
+        shutil.rmtree(dst, onerror=_on_rm_error)
+    shutil.copytree(src, dst, symlinks=symlinks, ignore=ignore)
+
+def _safe_copy(src: Path, dst: Path):
+    dst.parent.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(src, dst)
+
+def _cleanup_old_runs(max_age_hours=12):
+    try:
+        now = datetime.datetime.now().timestamp()
+        for run_dir in RUNS_DIR.iterdir():
+            try:
+                if not run_dir.is_dir():
+                    continue
+                mtime = run_dir.stat().st_mtime
+                age_h = (now - mtime) / 3600.0
+                if age_h > max_age_hours:
+                    shutil.rmtree(run_dir, onerror=_on_rm_error)
+            except Exception:
+                continue
+    except Exception:
+        pass
 
-    # Clean up old outputs
-    for item in OUTPUT_DIR.iterdir():
-        if item.is_dir():
-            shutil.rmtree(item)
+def _prepare_workspace(logs):
+    """Create isolated per-run workspace and copy needed code/assets into it."""
+    run_id = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") + "-" + uuid.uuid4().hex[:8]
+    work_dir = RUNS_DIR / run_id
+    work_dir.mkdir(parents=True, exist_ok=True)
+
+    # Per-run log & zip path
+    log_path = work_dir / "run.log"
+    zip_path = work_dir / "output.zip"
+
+    logs.append(f"🧩 New workspace: {work_dir.relative_to(ROOT)} (run_id={run_id})")
+
+    # Copy code/assets that do file IO so they are run-local (avoid shared writes)
+    # Keep copies as cheap as possible (symlinks=True when supported)
+    needed_dirs = ["posterbuilder", "Paper2Poster"]
+    for d in needed_dirs:
+        src = ROOT / d
+        if src.exists():
+            _copytree(src, work_dir / d, symlinks=True)
+            logs.append(f"   ↪ copied {d}/ → runs/{run_id}/{d}/ (symlink where possible)")
+
+    # template/ optional
+    tmpl = ROOT / "template"
+    if tmpl.exists():
+        _copytree(tmpl, work_dir / "template", symlinks=True)
+        logs.append("   ↪ copied template/")
+
+    # pipeline.py must live inside workspace so that ROOT_DIR=work_dir
+    _safe_copy(ROOT / "pipeline.py", work_dir / "pipeline.py")
+
+    # Create standard IO dirs in workspace
+    (work_dir / "input" / "pdf").mkdir(parents=True, exist_ok=True)
+    (work_dir / "input" / "logo").mkdir(parents=True, exist_ok=True)
+    (work_dir / "posterbuilder" / "latex_proj").mkdir(parents=True, exist_ok=True)
+
+    return run_id, work_dir, log_path, zip_path
+
+# ---------------------
+# Helpers for new features (post-processing)
+# ---------------------
+def _parse_rgb(s):
+    """Accepts '94,46,145' / '94 46 145' / '[94,46,145]' / '(94, 46, 145)' and returns (r,g,b) or None."""
+    if s is None:
+        return None
+    if isinstance(s, (tuple, list)) and len(s) == 3:
+        vals = s
+    else:
+        nums = re.findall(r"\d+", str(s))
+        if len(nums) < 3:
+            return None
+        vals = nums[:3]
+    try:
+        r, g, b = (int(vals[0]), int(vals[1]), int(vals[2]))
+        if any(v < 0 or v > 255 for v in (r, g, b)):
+            return None
+        return (r, g, b)
+    except Exception:
+        return None
+
+def _apply_meeting_logo(OUTPUT_DIR: Path, meeting_logo_file, logs):
+    """Replace output/poster_latex_proj/logos/right_logo.png if meeting_logo_file provided."""
+    if not meeting_logo_file:
+        return False
+
+    logos_dir = OUTPUT_DIR / "poster_latex_proj" / "logos"
+    target = logos_dir / "right_logo.png"
+    try:
+        logos_dir.mkdir(parents=True, exist_ok=True)
+        # Try to convert to PNG for safety
+        try:
+            from PIL import Image
+            img = Image.open(meeting_logo_file.name)
+            # preserve alpha if available
+            if img.mode not in ("RGB", "RGBA"):
+                img = img.convert("RGBA")
+            img.save(target, format="PNG")
+            logs.append(f"🖼️ Meeting logo converted to PNG and saved → {target.relative_to(OUTPUT_DIR)}")
+        except Exception as e:
+            # Fallback: raw copy with .png name
+            shutil.copy(meeting_logo_file.name, target)
+            logs.append(f"🖼️ Meeting logo copied (no conversion) → {target.relative_to(OUTPUT_DIR)} (note: ensure it's a valid PNG).")
+        return True
+    except Exception as e:
+        logs.append(f"⚠️ Failed to apply meeting logo: {e}")
+        return False
+
+def _apply_theme_rgb(OUTPUT_DIR: Path, rgb_tuple, logs):
+    """Replace \\definecolor{nipspurple}{RGB}{r,g,b} in poster_output.tex if rgb_tuple provided."""
+    if not rgb_tuple:
+        return False
+
+    tex_path = OUTPUT_DIR / "poster_latex_proj" / "poster_output.tex"
+    if not tex_path.exists():
+        logs.append(f"⚠️ Theme RGB skipped: {tex_path.relative_to(OUTPUT_DIR)} not found.")
+        return False
+
+    try:
+        content = tex_path.read_text(encoding="utf-8")
+        pattern = r"(\\definecolor\{nipspurple\}\{RGB\}\{)\s*\d+\s*,\s*\d+\s*,\s*\d+\s*(\})"
+        new_vals = f"{rgb_tuple[0]},{rgb_tuple[1]},{rgb_tuple[2]}"
+        new_content, n = re.subn(pattern, r"\1" + new_vals + r"\2", content, flags=re.MULTILINE)
+        if n > 0:
+            tex_path.write_text(new_content, encoding="utf-8")
+            logs.append(f"🎨 Theme color updated: themecolor = {{{new_vals}}} in {tex_path.relative_to(OUTPUT_DIR)}")
+            return True
         else:
-            item.unlink()
-    if ZIP_PATH.exists():
-        ZIP_PATH.unlink()
-    logs.append("🧹 Cleaned previous output.\n")
-    _write_logs(logs)
+            logs.append("⚠️ Theme RGB not applied: definecolor for 'themecolor' not found.")
+            return False
+    except Exception as e:
+        logs.append(f"⚠️ Failed to update theme RGB: {e}")
+        return False
+
+def _apply_left_logo(OUTPUT_DIR: Path, logo_files, logs):
+    """
+    Use the first institutional logo uploaded by the user:
+    - Copy it into output/poster_latex_proj/logos/ as left_logo.<ext>
+    - Replace 'logos/left_logo.png' in poster_output.tex with the proper file extension
+    Does NOT convert formats. Simply renames and rewrites the tex reference.
+    """
+    if not logo_files:
+        logs.append("ℹ️ No institutional logo uploaded.")
+        return False
+
+    # If multiple files component, take the first one
+    f = logo_files[0] if isinstance(logo_files, (list, tuple)) else logo_files
+    if not f:
+        logs.append("ℹ️ No institutional logo uploaded.")
+        return False
+
+    ext = Path(f.name).suffix or ".png"  # fallback to .png if no extension
+    logos_dir = OUTPUT_DIR / "poster_latex_proj" / "logos"
+    tex_path = OUTPUT_DIR / "poster_latex_proj" / "poster_output.tex"
+
+    try:
+        logos_dir.mkdir(parents=True, exist_ok=True)
+        dst = logos_dir / f"left_logo{ext}"
+        shutil.copy(f.name, dst)
+        logs.append(f"🏷️ Institutional logo copied to: {dst.relative_to(OUTPUT_DIR)}")
+    except Exception as e:
+        logs.append(f"⚠️ Failed to copy institutional logo: {e}")
+        return False
+
+    if not tex_path.exists():
+        logs.append("⚠️ poster_output.tex not found, cannot replace left_logo path.")
+        return False
+
+    try:
+        text = tex_path.read_text(encoding="utf-8")
+        old = "logos/left_logo.png"
+        new = f"logos/left_logo{ext}"
+
+        if old in text:
+            tex_path.write_text(text.replace(old, new), encoding="utf-8")
+            logs.append(f"🛠️ Replaced left_logo.png → left_logo{ext} in poster_output.tex")
+            return True
+
+        # Fallback (covers weird spacing or macro variations)
+        import re
+        pattern = r"(logos/left_logo)\.png"
+        new_text, n = re.subn(pattern, r"\1" + ext, text)
+
+        if n > 0:
+            tex_path.write_text(new_text, encoding="utf-8")
+            logs.append(f"🛠️ Replaced left_logo.png → left_logo{ext} (regex fallback)")
+            return True
+
+        logs.append("ℹ️ No left_logo.png reference found in poster_output.tex.")
+        return False
+
+    except Exception as e:
+        logs.append(f"⚠️ Failed to modify poster_output.tex: {e}")
+        return False
+
+
+# =====================
+# Gradio pipeline function (ISOLATED)
+# =====================
+def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files, meeting_logo_file, theme_rgb):
+    _cleanup_old_runs(RETENTION_HOURS)
+
+    start_time = datetime.datetime.now()
+    logs = [f"🚀 Starting pipeline at {_now_str()}"]
+
+    # --- Prepare per-run workspace ---
+    run_id, WORK_DIR, LOG_PATH, ZIP_PATH = _prepare_workspace(logs)
+    INPUT_DIR = WORK_DIR / "input"
+    OUTPUT_DIR = WORK_DIR / "output"
+    LOGO_DIR = INPUT_DIR / "logo"
+    POSTER_LATEX_DIR = WORK_DIR / "posterbuilder" / "latex_proj"
+
+    _write_logs(LOG_PATH, logs)
     yield "\n".join(logs), None
 
     # ====== Validation: must upload LOGO ======
-    # Gradio may return a single file or a list, normalize to list
     if logo_files is None:
         logo_files = []
     if not isinstance(logo_files, (list, tuple)):
         logo_files = [logo_files]
-    logo_files = [f for f in logo_files if f]  # filter None
+    logo_files = [f for f in logo_files if f]
 
     if len(logo_files) == 0:
         msg = "❌ You must upload at least one institutional logo (multiple allowed)."
         logs.append(msg)
-        _write_logs(logs)
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
         return
 
-    # Clear input/logo and then save new files
+    # Save logos into run-local dir
     for item in LOGO_DIR.iterdir():
         if item.is_file():
             item.unlink()
@@ -79,8 +315,8 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
         p = LOGO_DIR / Path(lf.name).name
         shutil.copy(lf.name, p)
         saved_logo_paths.append(p)
-    logs.append(f"🏷️ Saved {len(saved_logo_paths)} logo file(s) to: {LOGO_DIR}\n")
-    _write_logs(logs)
+    logs.append(f"🏷️ Saved {len(saved_logo_paths)} logo file(s) → {LOGO_DIR.relative_to(WORK_DIR)}")
+    _write_logs(LOG_PATH, logs)
     yield "\n".join(logs), None
 
     # ====== Handle uploaded PDF (optional) ======
@@ -90,46 +326,48 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
         pdf_dir.mkdir(parents=True, exist_ok=True)
         pdf_path = pdf_dir / Path(pdf_file.name).name
         shutil.copy(pdf_file.name, pdf_path)
-        logs.append(f"📄 Uploaded PDF saved to: {pdf_path}\n")
+        logs.append(f"📄 Uploaded PDF → {pdf_path.relative_to(WORK_DIR)}")
 
         # For pipeline Step 1.5 compatibility: also copy to input/paper.pdf
         canonical_pdf = INPUT_DIR / "paper.pdf"
         shutil.copy(pdf_file.name, canonical_pdf)
-        _write_logs(logs)
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
 
     # ====== Validate input source ======
     if not arxiv_url and not pdf_file:
         msg = "❌ Please provide either an arXiv link or upload a PDF file (choose one)."
         logs.append(msg)
-        _write_logs(logs)
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
         return
 
-    # ====== Build command ======
+    # ====== Build command (run INSIDE workspace) ======
     cmd = [
-        "python", "pipeline.py",
+        sys.executable, "pipeline.py",
         "--model_name_t", "gpt-5",
         "--model_name_v", "gpt-5",
         "--result_dir", "output",
         "--paper_latex_root", "input/latex_proj",
         "--openai_key", openai_key,
         "--gemini_key", "##",
-        "--logo_dir", str(LOGO_DIR)  # 👈 pass logo directory
+        "--logo_dir", str(LOGO_DIR)  # run-local logo dir
     ]
     if arxiv_url:
         cmd += ["--arxiv_url", arxiv_url]
-    # if pdf_path:
-    #     cmd += ["--pdf_path", str(pdf_path)]
+    # (Keep pdf via input/paper.pdf; pipeline will read it if exists)
 
-    # ====== Run command with REAL-TIME streaming ======
-    logs.append("\n======= REAL-TIME LOG =======\n")
-    _write_logs(logs)
+    logs.append("\n======= REAL-TIME LOG =======")
+    logs.append(f"cwd = runs/{WORK_DIR.name}")
+    logs.append(f"cmd = {' '.join(cmd)}")
+    _write_logs(LOG_PATH, logs)
     yield "\n".join(logs), None
 
+    # ====== Run with REAL-TIME streaming, inside workspace ======
     try:
         process = subprocess.Popen(
             cmd,
+            cwd=str(WORK_DIR),
             stdout=subprocess.PIPE,
             stderr=subprocess.STDOUT,
             text=True,
@@ -139,7 +377,7 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
     except Exception as e:
         msg = f"❌ Pipeline failed to start: {e}"
         logs.append(msg)
-        _write_logs(logs)
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
         return
 
@@ -148,24 +386,22 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
         while True:
             # Timeout guard
             if (datetime.datetime.now() - start_time).total_seconds() > TIMEOUT_SECONDS:
-                logs.append("❌ Pipeline timed out (30 min limit). Killing process...\n")
+                logs.append("❌ Pipeline timed out (30 min limit). Killing process…")
                 try:
                     process.kill()
                 except Exception:
                     pass
-                _write_logs(logs)
+                _write_logs(LOG_PATH, logs)
                 yield "\n".join(logs), None
                 return
 
             line = process.stdout.readline()
             if line:
-                # echo to HF Space logs as well
-                print(line, end="")
+                print(line, end="")  # echo to Space logs
                 logs.append(line.rstrip("\n"))
-                # write & occasionally yield to reduce UI churn
-                _write_logs(logs)
+                _write_logs(LOG_PATH, logs)
                 now = time.time()
-                if now - last_yield >= 0.3:  # throttle UI updates ~3/sec
+                if now - last_yield >= 0.3:
                     last_yield = now
                     yield "\n".join(logs), None
             elif process.poll() is not None:
@@ -174,19 +410,19 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
                 time.sleep(0.05)
 
         return_code = process.wait()
-        logs.append(f"\nProcess finished with code {return_code}\n")
-        _write_logs(logs)
+        logs.append(f"\nProcess finished with code {return_code}")
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
 
         if return_code != 0:
-            logs.append("❌ Process exited with non-zero status. See logs above.\n")
-            _write_logs(logs)
+            logs.append("❌ Process exited with non-zero status. See logs above.")
+            _write_logs(LOG_PATH, logs)
             yield "\n".join(logs), None
             return
 
     except Exception as e:
-        logs.append(f"❌ Error during streaming: {e}\n")
-        _write_logs(logs)
+        logs.append(f"❌ Error during streaming: {e}")
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
         return
     finally:
@@ -196,22 +432,44 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
         except Exception:
             pass
 
-    # ====== Check output & zip ======
+    # ====== Check output ======
     has_output = False
     try:
-        for _ in OUTPUT_DIR.iterdir():
-            has_output = True
-            break
+        if OUTPUT_DIR.exists():
+            for _ in OUTPUT_DIR.iterdir():
+                has_output = True
+                break
     except FileNotFoundError:
         has_output = False
 
     if not has_output:
         msg = "❌ No output generated. Please check logs above."
         logs.append(msg)
-        _write_logs(logs)
+        _write_logs(LOG_PATH, logs)
         yield "\n".join(logs), None
         return
 
+    # ====== NEW: Post-processing (optional features) ======
+    # 1) Optional meeting logo replacement
+    applied_logo = _apply_meeting_logo(OUTPUT_DIR, meeting_logo_file, logs)
+
+    # 2) Optional theme color update
+    rgb_tuple = _parse_rgb(theme_rgb)
+    if theme_rgb and not rgb_tuple:
+        logs.append(f"⚠️ Ignored Theme RGB input '{theme_rgb}': expected like '94,46,145'.")
+    applied_rgb = _apply_theme_rgb(OUTPUT_DIR, rgb_tuple, logs) if rgb_tuple else False
+
+    # 3) Optional institutional logo -> left_logo.<ext>
+    _apply_left_logo(OUTPUT_DIR, logo_files, logs)
+
+    _write_logs(LOG_PATH, logs)
+    yield "\n".join(logs), None
+
+
+    _write_logs(LOG_PATH, logs)
+    yield "\n".join(logs), None
+
+    # ====== Zip output (run-local) ======
     try:
         with zipfile.ZipFile(ZIP_PATH, 'w', zipfile.ZIP_DEFLATED) as zipf:
             for root, dirs, files in os.walk(OUTPUT_DIR):
@@ -219,43 +477,51 @@ def run_pipeline(arxiv_url, pdf_file, openai_key, logo_files):
                     file_path = Path(root) / file
                     arcname = file_path.relative_to(OUTPUT_DIR)
                     zipf.write(file_path, arcname=arcname)
-        logs.append(f"✅ Zipped output folder to {ZIP_PATH}\n")
+        logs.append(f"✅ Zipped output → {ZIP_PATH.relative_to(WORK_DIR)}")
     except Exception as e:
-        logs.append(f"❌ Failed to create zip: {e}\n")
+        logs.append(f"❌ Failed to create zip: {e}")
 
     end_time = datetime.datetime.now()
-    logs.append(f"🏁 Completed at {end_time.strftime('%Y-%m-%d %H:%M:%S')} (Duration: {(end_time - start_time).seconds}s)\n")
+    dur = (end_time - start_time).seconds
+    logs.append(f"🏁 Completed at {_now_str()} (Duration: {dur}s)")
+    logs.append(f"🆔 run_id = {WORK_DIR.name}")
 
-    _write_logs(logs)
+    _write_logs(LOG_PATH, logs)
     yield "\n".join(logs), (str(ZIP_PATH) if ZIP_PATH.exists() else None)
 
-def _write_logs(logs):
-    try:
-        with open(LOG_PATH, "w", encoding="utf-8") as f:
-            f.write("\n".join(logs))
-    except Exception:
-        # Avoid crashing UI if disk I/O fails
-        pass
-
-# ===================== Gradio UI =====================
+# =====================
+# Gradio UI
+# =====================
 iface = gr.Interface(
     fn=run_pipeline,
     inputs=[
         gr.Textbox(label="📘 ArXiv URL (choose one)", placeholder="https://arxiv.org/abs/2505.xxxxx"),
         gr.File(label="📄 Upload PDF (choose one)"),
         gr.Textbox(label="🔑 OpenAI API Key", placeholder="sk-...", type="password"),
-        gr.File(label="🏷️ Upload Institutional Logo(s) (required, multiple allowed)", file_count="multiple", file_types=["image"]),
+        gr.File(
+            label="🏷️ Institutional Logo (optional, multiple allowed)",
+            file_count="multiple",
+            file_types=["image"],
+        ),
+        gr.File(label="🧩 Optional: Conference Logo (replaces right_logo.png)", file_count="single", file_types=["image"]),
+        gr.Textbox(label="🎨 Optional: Theme RGB (e.g., 94,46,145)", placeholder="94,46,145"),
     ],
     outputs=[
-        gr.Textbox(label="🧾 Logs", lines=30, max_lines=50),
+        gr.Textbox(label="🧾 Logs (8~10 minutes)", lines=30, max_lines=50),
         gr.File(label="📦 Download Results (.zip)")
     ],
-    title="📄 Paper2Poster",
-    description=(
-        "Upload your paper, and the pipeline will automatically generate a fully compilable LaTeX poster; you can download the ZIP file and compile it yourself. Each paper takes approximately 6–10 minutes to process.\n"
-        "Provide either an arXiv link or upload a PDF file (choose one); the system will generate a poster and package it for download.\n"
-        "You must upload at least one institutional logo (multiple allowed).\n"
-    ),
+    title="🎓 Paper2Poster",
+    description="""
+    paper(https://arxiv.org/abs/2505.21497) | [GitHub](https://github.com/Paper2Poster/Paper2Poster) | project page (https://paper2poster.github.io/) 
+
+    # Paper2Poster
+
+    Upload a paper, generate a poster for you.  
+    Each paper takes approximately **8–10 minutes**.  
+
+    This work is based on the **[CAMEL-ai](https://camel-ai.org/)** framework.
+    """,
+
     allow_flagging="never",
 )
 

install_tectonic.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+set -e
+
+echo "📦 Installing tectonic..."
+
+if ! command -v tectonic &> /dev/null; then
+  wget -O /tmp/tectonic.tar.gz https://github.com/tectonic-typesetting/tectonic/releases/download/tectonic%400.15.0/tectonic-0.15.0-x86_64-unknown-linux-gnu.tar.gz
+  mkdir -p /tmp/tectonic
+  tar -xzf /tmp/tectonic.tar.gz -C /tmp/tectonic
+
+  # 找到可执行文件路径
+  TECTONIC_BIN=$(find /tmp/tectonic -type f -name tectonic | head -n 1)
+
+  # 默认安装路径
+  INSTALL_DIR="/usr/local/bin"
+
+  # 如果没有写入权限，就改到用户目录
+  if [ ! -w "$INSTALL_DIR" ]; then
+    INSTALL_DIR="$HOME/.local/bin"
+    mkdir -p "$INSTALL_DIR"
+    echo "⚠️  No permission for /usr/local/bin, installing to $INSTALL_DIR"
+  fi
+
+  cp "$TECTONIC_BIN" "$INSTALL_DIR/tectonic"
+  chmod +x "$INSTALL_DIR/tectonic"
+
+  # 自动提示 PATH 设置
+  if [[ ":$PATH:" != *":$INSTALL_DIR:"* ]]; then
+    echo "⚙️  You may need to add this to your ~/.bashrc:"
+    echo "export PATH=\$PATH:$INSTALL_DIR"
+  fi
+
+  echo "✅ Tectonic installed successfully at $INSTALL_DIR/tectonic"
+else
+  echo "Tectonic already installed."
+fi

install_tinytex.sh

@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+# ============================================================
+# install_tinytex_xelatex.sh
+# 一键安装 TinyTeX (用户级，无需 sudo)
+# 适用于 beamer / fontspec / gemini / cam 海报编译
+# ============================================================
+
+set -e  # 出错立即退出
+WORKDIR=$(pwd)
+echo "当前目录: $WORKDIR"
+
+# ------------------------------------------------------------
+# Step 0. 预清理旧残留
+# ------------------------------------------------------------
+echo "🧹 Step 0. 检查并清理旧 TinyTeX 安装或锁文件..."
+if [ -d "$HOME/.TinyTeX" ]; then
+  echo "⚠️ 检测到已有 ~/.TinyTeX 目录，尝试安全删除..."
+  # 杀掉占用 .nfs 文件的进程
+  PIDS=$(lsof +D "$HOME/.TinyTeX" 2>/dev/null | awk 'NR>1 {print $2}' | sort -u)
+  if [ -n "$PIDS" ]; then
+    echo "🔪 结束占用进程: $PIDS"
+    kill -9 $PIDS 2>/dev/null || true
+  fi
+  # 强制删除旧目录
+  rm -rf "$HOME/.TinyTeX" 2>/dev/null || true
+fi
+
+# 清除当前目录下可能存在的 texlive 安装残留
+rm -rf texlive install-tl* tinytex.profile 2>/dev/null || true
+
+# ------------------------------------------------------------
+# Step 1. 安装 TinyTeX
+# ------------------------------------------------------------
+echo "🚀 Step 1. 安装 TinyTeX 到用户目录 (~/.TinyTeX)..."
+wget -qO- "https://yihui.org/tinytex/install-unx.sh" | sh
+
+# ------------------------------------------------------------
+# Step 2. 设置 PATH 环境变量
+# ------------------------------------------------------------
+echo "✅ 安装完成，设置 PATH 环境变量..."
+export PATH=$HOME/.TinyTeX/bin/x86_64-linux:$PATH
+
+# 写入 ~/.bashrc 方便以后使用
+if ! grep -q ".TinyTeX/bin/x86_64-linux" ~/.bashrc; then
+  echo 'export PATH=$HOME/.TinyTeX/bin/x86_64-linux:$PATH' >> ~/.bashrc
+  echo "📝 已自动将 TinyTeX 路径写入 ~/.bashrc"
+fi
+
+# ------------------------------------------------------------
+# Step 3. 检查 XeLaTeX 是否可用
+# ------------------------------------------------------------
+echo "🧩 Step 2. 检查 XeLaTeX 版本..."
+if ! command -v xelatex >/dev/null 2>&1; then
+  echo "❌ XeLaTeX 未找到，安装可能失败，请检查 ~/.TinyTeX/bin 是否存在。"
+  exit 1
+else
+  xelatex --version | head -n 2
+fi
+
+# ------------------------------------------------------------
+# Step 4. 安装常用包
+# ------------------------------------------------------------
+echo "📦 Step 3. 安装必要的 LaTeX 包 (无 sudo)..."
+tlmgr option repository http://mirror.ctan.org/systems/texlive/tlnet
+tlmgr update --self --all --reinstall-forcibly-removed
+tlmgr install \
+  beamer beamerposter fontspec xcolor pgfplots anyfontsize type1cm \
+  ragged2e changepage geometry booktabs natbib tikz fp
+
+echo "✅ 所有包已安装。TinyTeX 可正常用于 XeLaTeX 海报编译。"
+echo "👉 现在你可以运行："
+echo "   xelatex poster_output.tex"

pipeline.py

@@ -21,6 +21,9 @@ P2P_ROOT   = ROOT_DIR / "Paper2Poster"
 PB_ROOT    = ROOT_DIR / "posterbuilder"
 sys.path.append(str(P2P_ROOT))
 
+print(f"🔒 Workspace ROOT_DIR = {ROOT_DIR}")
+print(f"🔒 This run is isolated under: {ROOT_DIR.resolve()}")
+
 def copy_folder(src_dir, dst_dir):
     src_dir = Path(src_dir)
     dst_dir = Path(dst_dir)
@@ -40,7 +43,7 @@ def safe_copy(src: Path, dst: Path):
     dst.parent.mkdir(parents=True, exist_ok=True)
     shutil.copy2(src, dst)
 
-def str2list(s): 
+def str2list(s):
     return [int(x) for x in s.split(',')]
 
 def run_paper2poster_content_build():
@@ -96,13 +99,7 @@ def _list_logo_files(logo_dir: Path):
     return files
 
 def _compose_logos_horizontally(logo_paths, out_path: Path, box_w=2000, box_h=476, gap=16):
-    """
-    宽度为硬约束：输出图像宽度必为 box_w（默认 2000px）。
-    多 logo 按比例统一缩放，拼接后刚好占满 box_w（包含间距）。
-    高度由比例自然决定，可能 < box_h，也可能 > box_h（甚至 > 2*box_h），不会再二次压缩。
-    透明背景，输出 PNG。
-    """
-    # 读取图片
+    # (same as your original; omitted comments for brevity)
     imgs = []
     for p in logo_paths:
         p = Path(p)
@@ -112,35 +109,27 @@ def _compose_logos_horizontally(logo_paths, out_path: Path, box_w=2000, box_h=47
     if n == 0:
         raise RuntimeError("No logo images found.")
 
-    # 原始总宽度（不含 gap）；拼接总宽 = sum(w_i) + gap*(n-1)
     widths  = [im.width for im in imgs]
     heights = [im.height for im in imgs]
     sum_w   = sum(widths)
     if sum_w <= 0:
         raise RuntimeError("All logo images have zero width.")
 
-    # 计算统一缩放比例，使：sum(w_i * s) + gap*(n-1) == box_w
-    # => s = (box_w - gap*(n-1)) / sum_w
     total_gap = max(0, gap * (n - 1))
     if box_w <= total_gap:
         raise ValueError(f"box_w({box_w}) too small vs total gaps({total_gap}). Increase box_w or reduce gap.")
     s = (box_w - total_gap) / float(sum_w)
 
-    # 按统一比例缩放（四舍五入到整数像素，避免累计误差）
     resized = []
     scaled_widths = []
-    scaled_heights = []
     for im, w, h in zip(imgs, widths, heights):
         nw = max(1, int(round(w * s)))
         nh = max(1, int(round(h * s)))
         resized.append(im.resize((nw, nh), Image.LANCZOS))
         scaled_widths.append(nw)
-        scaled_heights.append(nh)
 
-    # 由于整数取整，可能出现总宽 !=  box_w - total_gap；对若干图微调 1px 以精确对齐
     current_sum_w = sum(scaled_widths)
     diff = (box_w - total_gap) - current_sum_w
-    # 按从宽到窄/从大到小顺序均匀分配像素误差
     if diff != 0:
         order = sorted(range(n), key=lambda i: scaled_widths[i], reverse=(diff > 0))
         idx = 0
@@ -155,14 +144,13 @@ def _compose_logos_horizontally(logo_paths, out_path: Path, box_w=2000, box_h=47
                 remaining -= 1
             idx += 1
 
-    # 计算最终尺寸
     total_w = sum(scaled_widths) + total_gap
     assert total_w == box_w, f"width pack mismatch: got {total_w}, expect {box_w}"
     canvas_w = box_w
-    canvas_h = max(im.height for im in resized)  # 高度由比例自然决定（可能 > 2*box_h）
+    canvas_h = max(im.height for im in resized)
 
-    # 画布 & 居中摆放（垂直方向居中）
-    canvas = Image.new("RGBA", (canvas_w, canvas_h), (0, 0, 0, 0))
+    from PIL import Image as PILImage
+    canvas = PILImage.new("RGBA", (canvas_w, canvas_h), (0, 0, 0, 0))
     cur_x = 0
     for idx, im in enumerate(resized):
         y = (canvas_h - im.height) // 2
@@ -171,13 +159,8 @@ def _compose_logos_horizontally(logo_paths, out_path: Path, box_w=2000, box_h=47
         if idx != n - 1:
             cur_x += gap
 
-    # out_path.parent.mkdir(parents=True, exist_ok=True)
     canvas.save(out_path, format="PNG")
-    print(f"   🧩 Logos composed (width-locked) → {out_path.relative_to(ROOT_DIR)} "
-          f"(n={n}, final_size={canvas_w}x{canvas_h})")
-
-
-
+    print(f"   🧩 Logos composed (width-locked) → {out_path.relative_to(ROOT_DIR)} (n={n}, final_size={canvas_w}x{canvas_h})")
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(description='Paper2Video Generation Pipeline')
@@ -196,11 +179,11 @@ if __name__ == '__main__':
     args = parser.parse_args()
     print("start")
 
-    # ✅ 使用传入的 key 设置环境变量
+    # env
     os.environ["OPENAI_API_KEY"] = args.openai_key
     os.environ["GEMINI_API_KEY"] = args.gemini_key
 
-    # 清空 output
+    # clean and create run-local output
     output_dir = ROOT_DIR / "output"
     if output_dir.exists():
         print(f"   🧹 Clearing old output directory: {output_dir.relative_to(ROOT_DIR)}")
@@ -210,9 +193,7 @@ if __name__ == '__main__':
     (output_dir / "slide_imgs").mkdir(parents=True, exist_ok=True)
     print("   ✅ Created subfolders: latex_proj / poster_latex_proj / slide_imgs")
 
-    # ================
-    # Step 0: Download from arXiv
-    # ================
+    # Step 0: arXiv download (same as your original)
     try:
         if args.arxiv_url:
             import requests, tarfile
@@ -255,59 +236,13 @@ if __name__ == '__main__':
     except Exception as e:
         print(f"❌ Step 0 failed: {e}")
 
-    # =========================
-    # Step 1: Slide Generation
-    # =========================
-    # try:
-    #     print("🧩 Step 1: Generating Slides ...")
-    #     slide_latex_path = path.join(args.paper_latex_root, "slides.tex")
-    #     slide_image_dir = path.join(args.result_dir, 'slide_imgs')
-    #     os.makedirs(slide_image_dir, exist_ok=True)
-
-    #     start_time = time.time()
-    #     prompt_path = "./Paper2Video/src/prompts/slide_beamer_prompt.txt"
-
-    #     if args.if_tree_search:
-    #         usage_slide, beamer_path = latex_code_gen(
-    #             prompt_path=prompt_path,
-    #             tex_dir=args.paper_latex_root,
-    #             beamer_save_path=slide_latex_path,
-    #             model_config_ll=get_agent_config(args.model_name_t),
-    #             model_config_vl=get_agent_config(args.model_name_v),
-    #             beamer_temp_name=args.beamer_templete_prompt
-    #         )
-    #     else:
-    #         paper_latex_path = path.join(args.paper_latex_root, "main.tex")
-    #         usage_slide = latex_code_gen(
-    #             prompt_path=prompt_path,
-    #             tex_dir=args.paper_latex_root,
-    #             tex_path=paper_latex_path,
-    #             beamer_save_path=slide_latex_path,
-    #             model_config=get_agent_config(args.model_name_t)
-    #         )
-    #         beamer_path = slide_latex_path
-
-    #     if not os.path.exists(beamer_path):
-    #         raise FileNotFoundError(f"❌ Beamer PDF not found: {beamer_path}")
-
-    #     slide_imgs = convert_from_path(beamer_path, dpi=400)
-    #     for i, img in enumerate(slide_imgs):
-    #         img.save(path.join(slide_image_dir, f"{i+1}.png"))
-    #     print("✅ Step 1 done.")
-    # except Exception as e:
-    #     print(f"❌ Step 1 failed: {e}")
-
-    # =========================
-    # Step 1.5: Poster2Poster 内容生成
-    # =========================
+    # Step 1.5: content build
     try:
         run_paper2poster_content_build()
     except Exception as e:
         print(f"❌ Step 1.5 failed: {e}")
 
-    # =========================
-    # Step 2: Build Poster
-    # =========================
+    # Step 2: build poster
     try:
         print("🧩 Step 2: Building poster ...")
         build_poster()
@@ -315,9 +250,7 @@ if __name__ == '__main__':
     except Exception as e:
         print(f"❌ Step 2 failed: {e}")
 
-    # =========================
-    # Step 3: 导出 latex_proj & 处理 LOGO & 应用 template
-    # =========================
+    # Step 3: export latex & apply template & logos
     try:
         src_lp = PB_ROOT / "latex_proj"
         dst_lp = ROOT_DIR / "output" / "poster_latex_proj"
@@ -344,16 +277,13 @@ if __name__ == '__main__':
             print("⚠️ template directory not found, skipping Step 3.5.")
 
         logos_out_dir = dst_lp / "logos"
-        # logos_out_dir.mkdir(parents=True, exist_ok=True)
         left_logo_path = logos_out_dir / "left_logo.png"
 
         if len(logo_files) == 1:
-            # 单图：拷贝并转成 PNG（以确保一致）
             im = Image.open(logo_files[0]).convert("RGBA")
             im.save(left_logo_path, format="PNG")
             print(f"🖼️  Single logo saved → {left_logo_path.relative_to(ROOT_DIR)}")
         else:
-            # 多图：拼接
             _compose_logos_horizontally(logo_files, left_logo_path, box_w=2000, box_h=476, gap=16)
 
         print("✅ Step 3 done.")

posterbuilder/convert.py

@@ -85,12 +85,27 @@ def fix_latex_escaped_commands(s: str) -> str:
 def escape_text(s: str) -> str:
     if not s:
         return ""
+
+    # ✅ 保护 $...$ 内的内容
+    math = []
+    def save_math(m):
+        math.append(m.group(0))
+        return f"__MATH{len(math)-1}__"
+
+    s = re.sub(r"\${1,2}.*?\${1,2}", save_math, s)
+
     rep = {"&": r"\&", "%": r"\%", "$": r"\$", "#": r"\#", "_": r"\_",
            "{": r"\{", "}": r"\}", "~": r"\textasciitilde{}", "^": r"\textasciicircum{}"}
     for k, v in rep.items():
         s = s.replace(k, v)
+
+    # ✅ 恢复 math
+    for i, block in enumerate(math):
+        s = s.replace(f"__MATH{i}__", block)
+
     return s
 
+
 def soft_wrap_title_for_logo(title: str, first_limit=68, next_limit=72) -> str:
     if not title or len(title) <= first_limit: return title
     def break_at(s: str, limit: int):

posterbuilder/latex_proj/poster_output.tex

@@ -46,7 +46,7 @@
 % Title
 % ====================
 
-\title{Paper2Poster: Towards Multimodal Poster}
+\title{Paper2Poster: \ Towards Multimodal Poster Automation from Scientific Papers}
 
 \author{Wei Pang\textsuperscript{1}, Kevin Qinghong Lin\textsuperscript{2}, Xiangru Jian\textsuperscript{1}, Xi He\textsuperscript{1}, Philip Torr\textsuperscript{3}}
 
@@ -76,7 +76,7 @@
 
 
 % --- injected font tweaks ---
-\setbeamerfont{title}{size=\Huge}
+\setbeamerfont{title}{size=\huge}
 \setbeamerfont{author}{size=\Large}
 \setbeamerfont{institute}{size=\large}
 \setbeamerfont{block title}{size=\Large}
@@ -97,7 +97,7 @@
 \separatorcolumn
 \begin{column}{\colwidth}
 \begin{block}{Why Posters Are Hard}
-We target \textbf{single-page, multimodal compression} of \textit{20K+ tokens} into clear panels. Posters demand \textcolor{blue}{tight text–visual coupling}, \textbf{layout balance}, and \textit{readable density}. Pure LLM/VLM approaches \textcolor{red}{miss spatial feedback}, causing overflow and incoherence. We reveal that \textbf{visual-in-the-loop planning} is essential to preserve reading order, keep figures relevant, and sustain \textit{engagement} within hard space limits.
+We tackle \textbf{single-page multimodal compression}: dense papers must become legible posters with \textcolor{red}{tight spatial constraints}. Pure LLM or VLM approaches \textbf{struggle with layout}, missing \textit{reading order} and \textbf{overflow control}. We reveal \textcolor{blue}{visual-in-the-loop} planning is key to \textbf{clarity}, \textbf{balance}, and \textbf{engagement}.
 
 \begin{figure}
 \centering
@@ -106,8 +106,12 @@ We target \textbf{single-page, multimodal compression} of \textit{20K+ tokens} i
 
 \end{block}
 
-\begin{block}{Benchmark and Data}
-We launch the \textbf{Paper2Poster Benchmark}: \textcolor{blue}{100 paper–poster pairs} spanning \textit{280 topics}. Average input: \textcolor{blue}{20,370 tokens, 22.6 pages}. Output posters compress text by \textcolor{blue}{14.4×} and figures by \textcolor{blue}{2.6×}. Evaluation covers \textbf{Visual Quality}, \textbf{Textual Coherence}, \textbf{VLM-as-Judge}, and \textbf{PaperQuiz}. This suite spotlights \textit{semantic alignment}, \textbf{fluency}, and \textcolor{blue}{reader comprehension}.
+\begin{block}{Benchmark \& Task}
+We introduce \textbf{Paper2Poster} and the task: generate a \textbf{single-page}, well-balanced poster that faithfully conveys core ideas. The protocol measures \textit{what matters}: \textbf{visual alignment}, \textbf{text fluency}, \textbf{holistic quality}, and knowledge transfer via \textcolor{blue}{PaperQuiz}. Our setup \textbf{standardizes evaluation} for automated poster generation.
+\end{block}
+
+\begin{block}{Curated Diverse Dataset}
+Dataset spans \textcolor{blue}{100} paper–poster pairs (NeurIPS, ICML, ICLR). Papers average \textcolor{blue}{22.6} pages and \textcolor{blue}{20K+} tokens; posters average \textcolor{blue}{1.4K} tokens. We observe \textbf{14.4x} text compression and \textbf{2.6x} figure reduction. Coverage: CV (\textcolor{blue}{19\%}), NLP (\textcolor{blue}{17\%}), RL (\textcolor{blue}{10\%})—driving \textbf{robustness}.
 
 \begin{figure}
 \centering
@@ -116,8 +120,11 @@ We launch the \textbf{Paper2Poster Benchmark}: \textcolor{blue}{100 paper–post
 
 \end{block}
 
-\begin{block}{PaperQuiz: What Matters}
-We generate \textcolor{blue}{100 MCQs/paper}: \textbf{50 verbatim} + \textbf{50 interpretive}. Multiple VLM readers simulate \textit{novice-to-expert} audiences and answer from the poster only. Scores are length-penalized to reward \textbf{dense clarity}. Results \textbf{correlate with human judgment}, proving PaperQuiz captures \textcolor{blue}{information delivery} beyond surface visuals and discourages \textcolor{red}{verbose, unfocused designs}.
+\end{column}
+\separatorcolumn
+\begin{column}{\colwidth}
+\begin{block}{Four-Pronged Evaluation}
+Our \textbf{four-pronged} suite tests end-to-end quality: Visual Quality via \textcolor{blue}{AltCLIP} similarity and \textbf{figure relevance}; Textual Coherence via \textcolor{blue}{PPL} (Llama-2-7B); VLM-as-Judge across \textbf{6 criteria}; and \textcolor{blue}{PaperQuiz} with length-aware penalties rewarding \textbf{dense, readable} designs.
 
 \begin{figure}
 \centering
@@ -126,11 +133,8 @@ We generate \textcolor{blue}{100 MCQs/paper}: \textbf{50 verbatim} + \textbf{50
 
 \end{block}
 
-\end{column}
-\separatorcolumn
-\begin{column}{\colwidth}
 \begin{block}{PosterAgent Pipeline}
-Our \textbf{top-down, visual-in-the-loop} agent compresses long papers into coherent posters. • \textbf{Parser} builds a structured asset library. • \textbf{Planner} aligns text–visual pairs and produces a \textcolor{blue}{binary-tree layout}. • \textbf{Painter–Commenter} renders panels via code and uses VLM feedback to fix \textcolor{red}{overflow} and misalignment. The result: \textbf{balanced, legible}, editable posters.
+PosterAgent is \textbf{top-down, visual-in-the-loop}. \textit{Parser} builds a semantic asset library; \textit{Planner} aligns text–visual pairs and uses \textcolor{blue}{binary-tree} layouts to preserve \textbf{reading order}. \textit{Painter-Commenter} renders panels, applies \textcolor{blue}{zoom-in} VLM feedback, and fixes \textbf{overflow} and \textbf{alignment}—yielding concise, coherent posters.
 
 \begin{figure}
 \centering
@@ -139,16 +143,12 @@ Our \textbf{top-down, visual-in-the-loop} agent compresses long papers into cohe
 
 \end{block}
 
-\begin{block}{Parser: Structured Assets}
-We distill PDFs into \textbf{section synopses} and \textit{figure/table assets} using \textcolor{blue}{MARKER} and \textcolor{blue}{DOCLING}, then LLM summarization. The asset library preserves \textbf{hierarchy} and \textit{semantics} while shrinking context for efficient planning. This step boosts \textbf{visual-semantic matching} and reduces \textcolor{red}{noise}, enabling reliable downstream \textit{layout reasoning}.
-\end{block}
-
-\begin{block}{Planner: Layout Mastery}
-We semantically match \textbf{sections ↔ figures} and allocate space via a \textcolor{blue}{binary-tree layout} that preserves \textit{reading order}, aspect ratios, and \textbf{content length} estimates. Panels are populated iteratively, ensuring \textbf{text brevity} and \textit{visual balance}. This strategy stabilizes coordinates and avoids \textcolor{red}{LLM numeric drift} in absolute placements.
+\begin{block}{Main Results}
+Across metrics, \textbf{PosterAgent} variants beat multi-agent baselines. We attain \textcolor{blue}{state-leading figure relevance} and near-\textbf{human} visual similarity. GPT-4o pixel posters look good but show \textcolor{red}{noisy text} and high \textcolor{red}{PPL}. VLM-as-Judge scores place PosterAgent-4o at \textcolor{blue}{3.72} overall, approaching GT posters.
 
 \begin{figure}
 \centering
-\includegraphics[width=0.80\linewidth]{figures/paper-picture-30.png}
+\includegraphics[width=0.80\linewidth]{figures/paper-table-1.png}
 \end{figure}
 
 \end{block}
@@ -156,30 +156,26 @@ We semantically match \textbf{sections ↔ figures} and allocate space via a \te
 \end{column}
 \separatorcolumn
 \begin{column}{\colwidth}
-\begin{block}{Painter–Commenter Loop}
-The \textbf{Painter} turns section–figure pairs into crisp bullets and executable \textcolor{blue}{python-pptx} code, rendering draft panels. The \textbf{Commenter} VLM zooms into panels, using \textit{in-context examples} to flag \textcolor{red}{overflow} or \textcolor{red}{blankness}. Iterations continue until \textbf{fit and alignment} are achieved, producing \textit{readable, compact} panels with minimal revision cycles.
+\begin{block}{PaperQuiz Insights}
+\textcolor{blue}{PaperQuiz} tracks human judgment and rewards \textbf{informative brevity}. With penalties, GT posters lead; \textbf{PosterAgent} tops automated methods. Open-source \textcolor{blue}{Qwen-2.5} stacks stay \textbf{competitive}. Stronger reader VLMs exploit \textbf{structured layouts}, outperforming blog-like or \textcolor{red}{text-garbling} image generations.
 
 \begin{figure}
 \centering
-\includegraphics[width=0.80\linewidth]{figures/paper-picture-61.png}
+\includegraphics[width=0.80\linewidth]{figures/paper-picture-9.png}
 \end{figure}
 
 \end{block}
 
-\begin{block}{Results: Stronger, Leaner}
-Our open-source variants beat \textcolor{blue}{4o-driven multi-agents} on most metrics, with \textcolor{blue}{87\% fewer tokens}. We hit \textbf{state-of-the-art figure relevance}, near-\textit{GT} visual similarity, and \textbf{high VLM-as-Judge} scores. PaperQuiz confirms \textbf{better knowledge transfer}. Cost is tiny: \textcolor{blue}{\$0.0045–\$0.55/poster}. Key bottleneck remains \textcolor{red}{Engagement}, guiding future design.
+\begin{block}{Efficient, Open, Scalable}
+Our pipeline slashes tokens by \textcolor{blue}{60–87\%}. PosterAgent-4o uses \textcolor{blue}{101K} tokens (\textcolor{blue}{\$0.55}); PosterAgent-Qwen uses \textcolor{blue}{47.6K} (\textcolor{blue}{\$0.0045}). Runtime ≈ \textcolor{blue}{4.5 min}. \textcolor{red}{Bottleneck}: sequential panel refinement; \textbf{future} parallelism, external knowledge, and human-in-the-loop will boost \textbf{engagement}.
 
 \begin{figure}
 \centering
-\includegraphics[width=0.80\linewidth]{figures/paper-table-1.png}
+\includegraphics[width=0.80\linewidth]{figures/paper-table-8.png}
 \end{figure}
 
 \end{block}
 
-\begin{block}{Limits and Next Steps}
-Current bottleneck: \textbf{sequential panel refinement} slows throughput (\textasciitilde{}\textcolor{blue}{4.5 min/doc}). We plan \textbf{panel-level parallelism}, \textit{external knowledge} integration (e.g., OpenReview), and \textbf{human-in-the-loop} editing for higher \textcolor{blue}{engagement}. These upgrades aim to boost \textbf{runtime, interactivity}, and \textit{visual storytelling}, pushing toward fully automated \textbf{author-grade posters}.
-\end{block}
-
 \end{column}
 \separatorcolumn
 \end{columns}

requirements.txt

@@ -68,6 +68,7 @@ marko==2.1.2
 python-docx==1.1.2
 deepsearch-glm==1.0.0
 openpyxl==3.1.5
+tectonic @ https://github.com/tectonic-typesetting/tectonic/releases/download/tectonic%400.15.0/tectonic-0.15.0-x86_64-unknown-linux-gnu.tar.gz
 
 # ========= LangChain / LLMs =========
 langchain==0.3.17