Skip to content

K-Dense-AI/scientific-agent-skills / skills/scientific-slides

scientific-slides

Verified

Builds slide decks and presentations for research talks. Used for making PowerPoint slides, conference presentations, seminar talks, research presentations, thesis defense slides, or any scientific talk. Provides slide structure, design templates, timing guidance, and visual validation. Works with PowerPoint and LaTeX Beamer.

scripts/generate_slide_image_ai.py

View on GitHub
#!/usr/bin/env python3
"""
AI-powered slide image generation using Nano Banana 2.

This script generates presentation slides or slide visuals using AI:
- full_slide mode: Generate complete slides with title, content, and visuals (for PDF workflow)
- visual_only mode: Generate just images/figures to place on slides (for PPT workflow)

Supports attaching conceptual or style references. Attachments may be redrawn;
embed original quantitative figures with native slide composition instead.

Uses smart iterative refinement:
1. Generate initial image with Nano Banana 2
2. Quality review using Gemini 3.7 Flash
3. Only regenerate if quality is below threshold
4. Repeat until quality meets standards (max iterations)

Requirements:
    - OPENROUTER_API_KEY environment variable
    - requests library

Usage:
    # Full slide for PDF workflow
    python generate_slide_image_ai.py "Title: Introduction to ML\nKey points: supervised learning, neural networks" -o slide_01.png
    
    # Visual only for PPT workflow
    python generate_slide_image_ai.py "Neural network architecture diagram" -o figure.png --visual-only
    
    # With reference images attached
    python generate_slide_image_ai.py "Draft a conceptual overview in this style" -o slide.png --attach style_reference.png
"""

import argparse
import base64
import json
import os
import re
import sys
import time
from pathlib import Path
from typing import Optional, Dict, Any, List, NamedTuple


try:
    import requests
except ImportError:
    print("Error: requests library not found. Install with: uv pip install requests")
    sys.exit(1)


class ReviewResult(NamedTuple):
    """The outcome of one quality review.

    A named tuple rather than a bare tuple because the review has several
    failure modes -- the model returns no choices, the request raises, the
    answer arrives in an unexpected shape -- and each of them used to be a
    chance to return the wrong number of values to the caller.

    `score` is None whenever no usable number was parsed, and `reviewed` is
    False whenever the review never produced an answer at all. Neither case
    invents a score: a slide whose quality was not measured is reported as
    unmeasured, not as passing.
    """

    critique: str
    score: Optional[float]
    needs_improvement: bool
    reviewed: bool
    error: Optional[str] = None


# Markdown decoration the reviewer routinely wraps its answer in. Gemini writes
# "**SCORE:** 8.5" far more often than the bare "SCORE: 8.5" the rubric asks
# for, and a pattern that does not allow for it silently scores every image
# at whatever the default happens to be.
_MARKUP = re.compile(r"[*_`#]")

_SCORE_PATTERN = re.compile(r"\bSCORE\s*:?\s*(\d+(?:\.\d+)?)", re.IGNORECASE)
_LOOSE_SCORE_PATTERN = re.compile(
    r"(?:score|rating|quality)\s*[:\s]\s*(\d+(?:\.\d+)?)\s*(?:/\s*10)?", re.IGNORECASE
)
# The underscore is optional because normalization strips it out of
# NEEDS_IMPROVEMENT along with the surrounding markdown.
_VERDICT_PATTERN = re.compile(
    r"^\s*VERDICT\s*:?\s*(ACCEPTABLE|NEEDS[_ ]?IMPROVEMENT)", re.IGNORECASE | re.MULTILINE
)


def _normalize_review_text(text: str) -> str:
    """Strip markdown decoration so the rubric's fields can be found."""
    return _MARKUP.sub("", text or "")


def _parse_score(text: str) -> Optional[float]:
    """Return the reviewer's 0-10 score, or None when there isn't a usable one.

    None rather than a default: a fabricated score is indistinguishable from a
    real one once it reaches the review log, and a default near the middle of
    the range quietly passes lenient thresholds while failing strict ones.
    """
    normalized = _normalize_review_text(text)
    match = _SCORE_PATTERN.search(normalized) or _LOOSE_SCORE_PATTERN.search(normalized)
    if not match:
        return None
    score = float(match.group(1))
    # A number outside the rubric's range means something other than the score
    # was matched -- a slide number, a percentage, an iteration number.
    return score if 0.0 <= score <= 10.0 else None


def _parse_verdict(text: str) -> Optional[str]:
    """Return ACCEPTABLE / NEEDS_IMPROVEMENT from the verdict line, or None.

    Anchored to the start of a line. The rubric prompt spells out both tokens
    when it explains the response format, so a reviewer that echoes those
    instructions back would fail a perfectly good image under a plain
    substring search.
    """
    match = _VERDICT_PATTERN.search(_normalize_review_text(text))
    if not match:
        return None
    verdict = match.group(1).upper().replace(" ", "").replace("_", "")
    return "NEEDS_IMPROVEMENT" if verdict == "NEEDSIMPROVEMENT" else "ACCEPTABLE"


def _resolve_api_key(explicit: Optional[str] = None) -> Optional[str]:
    """Resolve the OpenRouter key from --api-key, the environment, then any .env file.

    The .env scan walks up from the working directory and finally checks the
    script's own directory, so running from anywhere inside a project picks up
    the key at its root. Only the standard library is used: python-dotenv is a
    common omission, and a missing optional dependency should not read as a
    missing credential.
    """
    if explicit:
        return explicit

    from_env = os.environ.get("OPENROUTER_API_KEY", "").strip()
    if from_env:
        return from_env

    cwd = Path.cwd()
    for directory in [cwd, *cwd.parents, Path(__file__).resolve().parent]:
        env_file = directory / ".env"
        if not env_file.is_file():
            continue
        try:
            content = env_file.read_text(encoding="utf-8", errors="replace")
        except OSError:
            continue
        for raw in content.splitlines():
            line = raw.strip()
            if line.startswith("#") or "=" not in line:
                continue
            name, _, value = line.partition("=")
            if name.strip() == "OPENROUTER_API_KEY":
                value = value.strip().strip('"').strip("'")
                if value:
                    return value

    return None


class SlideImageGenerator:
    """Generate presentation slides or visuals using AI with iterative refinement.
    
    Two modes:
    - full_slide: Generate complete slide with title, content, visuals (for PDF workflow)
    - visual_only: Generate just the image/figure for a slide (for PPT workflow)
    """
    
    # Quality threshold for presentations (lower than journal/conference papers)
    QUALITY_THRESHOLD = 6.5
    
    # Guidelines for generating full slides (complete slide images)
    FULL_SLIDE_GUIDELINES = """
Create a professional presentation slide image with these requirements:

SLIDE LAYOUT (16:9 aspect ratio):
- Clean, modern slide design
- Clear visual hierarchy: title at top, content below
- Generous margins (at least 5% on all sides)
- Balanced composition with intentional white space

TYPOGRAPHY:
- LARGE, bold title text (easily readable from distance)
- Clear, sans-serif fonts throughout
- High contrast text (dark on light or light on dark)
- Bullet points or key phrases, NOT paragraphs
- Maximum 5-6 lines of text content
- Use the supplied author/presenter; omit it when no name is provided

VISUAL ELEMENTS:
- Use GENERIC, simple images and icons - avoid overly specific or detailed imagery
- MINIMAL extra elements - no decorative borders, shadows, or flourishes
- Visuals should support and enhance the message, not distract
- Professional, clean aesthetic with restraint
- Consistent color scheme (2-3 main colors only)
- Prefer abstract/conceptual visuals over literal representations

PROFESSIONAL MINIMALISM:
- Less is more: favor empty space over additional elements
- No unnecessary decorations, gradients, or visual noise
- Clean lines and simple shapes
- Focused content without visual clutter
- Corporate/academic level of professionalism

PRESENTATION QUALITY:
- Designed for projection (high contrast)
- Bold, impactful design that commands attention
- Professional and polished appearance
- No cluttered or busy layouts
- Consistent styling throughout the deck
"""

    # Guidelines for generating slide visuals only (figures/images for PPT)
    VISUAL_ONLY_GUIDELINES = """
Create a high-quality visual/figure for a presentation slide:

IMAGE QUALITY:
- Clean, professional appearance
- High resolution and sharp details
- Suitable for embedding in a slide

DESIGN:
- Simple, clear composition with MINIMAL elements
- High contrast for projection readability
- No text unless essential to the visual
- Transparent or white background preferred
- GENERIC imagery - avoid overly specific or detailed visuals

PROFESSIONAL MINIMALISM:
- Favor simplicity over complexity
- No decorative elements, shadows, or flourishes
- Clean lines and simple shapes only
- Remove any unnecessary visual noise
- Abstract/conceptual rather than literal representations

STYLE:
- Modern, professional aesthetic
- Colorblind-friendly colors
- Bold but restrained imagery
- Suitable for scientific/professional presentations
- Corporate/academic level of polish
"""
    
    def __init__(self, api_key: Optional[str] = None, verbose: bool = False):
        """
        Initialize the generator.
        
        Args:
            api_key: OpenRouter API key (or use OPENROUTER_API_KEY env var)
            verbose: Print detailed progress information
        """
        self.api_key = _resolve_api_key(api_key)

        if not self.api_key:
            raise ValueError(
                "OPENROUTER_API_KEY not found. Please either:\n"
                "  1. Set the OPENROUTER_API_KEY environment variable\n"
                "  2. Add OPENROUTER_API_KEY to your .env file\n"
                "  3. Pass api_key parameter to the constructor\n"
                "Get your API key from: https://openrouter.ai/keys"
            )
        
        self.verbose = verbose
        self._last_error = None
        self.base_url = "https://openrouter.ai/api/v1"
        # Nano Banana 2 for image generation. The slug must be an image-output
        # model; a text-only chat model is rejected with "No endpoints found that
        # support the requested output modalities".
        # https://openrouter.ai/google/gemini-3.1-flash-image
        self.image_model = "google/gemini-3.1-flash-image"
        # Gemini 3.7 Flash for quality review - reads the image, answers in text
        self.review_model = "google/gemini-3.7-flash"
        
    def _log(self, message: str):
        """Log message if verbose mode is enabled."""
        if self.verbose:
            print(f"[{time.strftime('%H:%M:%S')}] {message}")
    
    def _post_request(self, endpoint: str, payload: Dict[str, Any]) -> Dict[str, Any]:
        """Send one non-streaming request; never automatically replay a paid call."""
        headers = {
            "Authorization": f"Bearer {self.api_key}",
            "Content-Type": "application/json",
            "X-Title": "Scientific Slide Generator",
        }
        self._log(f"POST {endpoint} using {payload['model']}")
        try:
            response = requests.post(
                f"{self.base_url}/{endpoint}", headers=headers,
                json=payload, timeout=120,
            )
            try:
                response_json = response.json()
            except ValueError:
                raise RuntimeError(
                    f"API returned non-JSON data (HTTP {response.status_code})"
                ) from None
            if not isinstance(response_json, dict):
                raise RuntimeError("API response must be a JSON object")
            if response.status_code != 200 or "error" in response_json:
                error = response_json.get("error", {})
                detail = error.get("message", "request rejected") if isinstance(error, dict) else str(error)
                # Providers occasionally echo input; do not dump whole bodies or credentials.
                detail = str(detail).replace(self.api_key, "[redacted]")[:500]
                raise RuntimeError(f"API request failed (HTTP {response.status_code}): {detail}")
            return response_json
        except requests.exceptions.Timeout:
            raise RuntimeError("API request timed out after 120 seconds") from None
        except requests.exceptions.RequestException as exc:
            detail = str(exc).replace(self.api_key, "[redacted]")
            raise RuntimeError(f"API request failed: {detail}") from None

    def _make_request(self, model: str, messages: List[Dict[str, Any]],
                      modalities: Optional[List[str]] = None) -> Dict[str, Any]:
        """Use chat completions for the text-and-image vision review."""
        payload = {"model": model, "messages": messages}
        if modalities:
            payload["modalities"] = modalities
        return self._post_request("chat/completions", payload)

    def _extract_image_from_response(self, response: Dict[str, Any]) -> bytes:
        """Decode the dedicated Image API response and enforce this helper's PNG contract.

        The provider may omit media_type; the PNG signature still has to match.
        This is a format check, not a full image-decoder or scientific-content check.
        """
        data = response.get("data")
        if not isinstance(data, list) or not data or not isinstance(data[0], dict):
            raise RuntimeError("Image API returned no data[0] image")
        item = data[0]
        if item.get("media_type") not in (None, "image/png"):
            raise RuntimeError(f"Expected PNG output; received {item['media_type']}")
        encoded = item.get("b64_json")
        if not isinstance(encoded, str) or not encoded:
            raise RuntimeError("Image API returned no b64_json image")
        try:
            image_data = base64.b64decode(encoded, validate=True)
        except ValueError:
            raise RuntimeError("Image API returned invalid base64 image data") from None
        if not image_data.startswith(b"\x89PNG\r\n\x1a\n"):
            raise RuntimeError("Image API output does not have a PNG signature")
        return image_data

    def _image_to_base64(self, image_path: str) -> str:
        """Convert image file to base64 data URL."""
        with open(image_path, "rb") as f:
            image_data = f.read()
        
        ext = Path(image_path).suffix.lower()
        mime_types = {
            ".png": "image/png",
            ".jpg": "image/jpeg",
            ".jpeg": "image/jpeg",
            ".gif": "image/gif",
            ".webp": "image/webp"
        }
        if ext not in mime_types:
            raise ValueError(f"Unsupported attachment format: {ext}; use PNG, JPEG, GIF, or WebP")
        mime_type = mime_types[ext]
        
        base64_data = base64.b64encode(image_data).decode("utf-8")
        return f"data:{mime_type};base64,{base64_data}"
    
    def generate_image(self, prompt: str, attachments: Optional[List[str]] = None) -> Optional[bytes]:
        """Generate one widescreen PNG using the dedicated Image API.

        References guide generation; they are not preserved as original pixels.
        A missing or unsupported reference aborts before a paid request.
        """
        self._last_error = None
        try:
            if attachments and len(attachments) > 14:
                raise ValueError("This image model accepts at most 14 reference images")
            payload = {
                "model": self.image_model, "prompt": prompt,
                "n": 1, "aspect_ratio": "16:9",
            }
            if attachments:
                payload["input_references"] = [
                    {"type": "image_url", "image_url": {"url": self._image_to_base64(path)}}
                    for path in attachments
                ]
            # output_format is not in this model's endpoint capability map.
            # Validate the returned PNG bytes instead of sending unsupported knobs.
            response = self._post_request("images", payload)
            image_data = self._extract_image_from_response(response)
            self._log(f"[OK] Generated PNG ({len(image_data)} bytes)")
            return image_data
        except (RuntimeError, ValueError, OSError) as exc:
            self._last_error = str(exc)
            self._log(f"[FAIL] Generation failed: {self._last_error}")
            return None

    def review_image(self, image_path: str, original_prompt: str, 
                    iteration: int, visual_only: bool = False,
                    max_iterations: int = 2) -> ReviewResult:
        """Review generated image using Gemini 3.7 Flash."""
        image_data_url = self._image_to_base64(image_path)
        threshold = self.QUALITY_THRESHOLD
        
        image_type = "slide visual/figure" if visual_only else "presentation slide"
        
        review_prompt = f"""You are an expert reviewer evaluating a {image_type} for presentation quality.

ORIGINAL REQUEST: {original_prompt}

QUALITY THRESHOLD: {threshold}/10
ITERATION: {iteration}/{max_iterations}

Evaluate this {image_type} on these criteria:

1. **Visual Impact** (0-2 points)
   - Bold, attention-grabbing design
   - Professional appearance
   - Suitable for projection

2. **Clarity** (0-2 points)
   - Easy to understand at a glance
   - Clear visual hierarchy
   - Not cluttered or busy

3. **Readability** (0-2 points)
   - Text is large and readable (if present)
   - High contrast
   - Clean typography

4. **Composition** (0-2 points)
   - Balanced layout
   - Good use of space
   - Appropriate margins

5. **Relevance** (0-2 points)
   - Matches the requested content
   - Appropriate style for presentations
   - Professional quality

RESPOND IN THIS EXACT FORMAT:
SCORE: [total score 0-10]

STRENGTHS:
- [strength 1]
- [strength 2]

ISSUES:
- [issue 1 if any]
- [issue 2 if any]

VERDICT: [ACCEPTABLE or NEEDS_IMPROVEMENT]

If score >= {threshold}, the image is ACCEPTABLE.
If score < {threshold}, mark as NEEDS_IMPROVEMENT with specific suggestions."""

        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": review_prompt},
                    {"type": "image_url", "image_url": {"url": image_data_url}}
                ]
            }
        ]
        
        try:
            response = self._make_request(model=self.review_model, messages=messages)
            
            choices = response.get("choices", [])
            if not choices:
                # A normal enough outcome -- a rate limit, a content filter, a
                # provider hiccup. The image itself is fine; only its review is
                # missing, and saying so beats inventing a score.
                reason = "the review model returned no choices"
                self._log(f"[WARN] Review unavailable: {reason}")
                return ReviewResult(
                    f"Review unavailable: {reason}.",
                    None, False, reviewed=False, error=reason,
                )

            message = choices[0].get("message", {})
            content = message.get("content", "")

            reasoning = message.get("reasoning", "")
            if reasoning and not content:
                content = reasoning

            if isinstance(content, list):
                text_parts = []
                for block in content:
                    if isinstance(block, dict) and block.get("type") == "text":
                        text_parts.append(block.get("text", ""))
                content = "\n".join(text_parts)

            score = _parse_score(content)
            verdict = _parse_verdict(content)

            if score is None and verdict is None:
                reason = "no score or verdict found in the review"
                self._log(f"[WARN] Review unusable: {reason}")
                return ReviewResult(
                    content if content else f"Review unusable: {reason}.",
                    None, False, reviewed=True, error=reason,
                )

            needs_improvement = verdict == "NEEDS_IMPROVEMENT" or (
                score is not None and score < threshold
            )

            shown = f"{score}/10" if score is not None else "not stated"
            self._log(f"[OK] Review complete (Score: {shown}, Threshold: {threshold}/10)")

            return ReviewResult(
                content if content else "Review returned no text",
                score,
                needs_improvement,
                reviewed=True,
                error=None if score is not None else "no score found in the review",
            )
        except Exception as e:
            # A failed review must not fail the run -- the image is already
            # generated -- but it must not read as a pass either.
            self._log(f"[WARN] Review failed: {str(e)}")
            return ReviewResult(
                f"Review failed: {str(e)}",
                None, False, reviewed=False, error=str(e),
            )
    
    def improve_prompt(self, original_prompt: str, critique: str, 
                      iteration: int, visual_only: bool = False) -> str:
        """Improve the generation prompt based on critique."""
        guidelines = self.VISUAL_ONLY_GUIDELINES if visual_only else self.FULL_SLIDE_GUIDELINES
        
        return f"""{guidelines}

USER REQUEST: {original_prompt}

ITERATION {iteration}: Based on previous feedback, address these specific improvements:
{critique}

Generate an improved version that addresses all the critique points."""
    
    def generate_slide(self, user_prompt: str, output_path: str,
                      visual_only: bool = False,
                      iterations: int = 2,
                      attachments: Optional[List[str]] = None) -> Dict[str, Any]:
        """
        Generate a slide image or visual with iterative refinement.
        
        Args:
            user_prompt: Description of the slide/visual to generate
            output_path: Path to save final image
            visual_only: If True, generate just the visual (for PPT workflow)
            iterations: Maximum refinement iterations (default: 2)
            attachments: Optional list of image file paths to attach as context
            
        Returns:
            Dictionary with generation results and metadata
        """
        if isinstance(iterations, bool) or not isinstance(iterations, int) or not 1 <= iterations <= 2:
            raise ValueError("iterations must be 1 or 2")
        output_path = Path(output_path)
        if output_path.suffix.lower() != ".png":
            raise ValueError("Output path must end in .png; format conversion is not supported")
        output_dir = output_path.parent
        output_dir.mkdir(parents=True, exist_ok=True)
        
        base_name = output_path.stem
        extension = output_path.suffix or ".png"
        
        mode = "visual_only" if visual_only else "full_slide"
        guidelines = self.VISUAL_ONLY_GUIDELINES if visual_only else self.FULL_SLIDE_GUIDELINES
        
        results = {
            "user_prompt": user_prompt,
            "mode": mode,
            "quality_threshold": self.QUALITY_THRESHOLD,
            "image_model": self.image_model,
            "review_model": self.review_model,
            "quality_met": False,
            "termination_reason": None,
            "attachments": attachments or [],
            "iterations": [],
            "final_image": None,
            # None, not 0.0 -- a run whose review never completed has no score,
            # and a number here would be read as one the reviewer gave.
            "final_score": None,
            "final_reviewed": False,
            "success": False,
            "early_stop": False
        }
        
        current_prompt = f"""{guidelines}

USER REQUEST: {user_prompt}

Generate a high-quality {'visual/figure' if visual_only else 'presentation slide'} that meets all the guidelines above."""
        
        print(f"\n{'='*60}")
        print(f"Generating Slide {'Visual' if visual_only else 'Image'}")
        print(f"{'='*60}")
        print(f"Description: {user_prompt[:100]}{'...' if len(user_prompt) > 100 else ''}")
        print(f"Mode: {mode}")
        if attachments:
            print(f"Attachments: {len(attachments)} image(s)")
            for att in attachments:
                print(f"  - {att}")
        print(f"Quality Threshold: {self.QUALITY_THRESHOLD}/10")
        print(f"Max Iterations: {iterations}")
        print(f"Output: {output_path}")
        print(f"{'='*60}\n")
        
        # Track temporary files for cleanup
        temp_files = []
        final_image_data = None
        
        for i in range(1, iterations + 1):
            print(f"\n[Iteration {i}/{iterations}]")
            print("-" * 40)
            
            print(f"Generating image with Nano Banana 2...")
            image_data = self.generate_image(current_prompt, attachments=attachments)
            
            if not image_data:
                error_msg = self._last_error or 'Image generation failed'
                print(f"[FAIL] Generation failed: {error_msg}")
                results["iterations"].append({
                    "iteration": i,
                    "success": False,
                    "error": error_msg
                })
                results["termination_reason"] = "generation_failed"
                break
            
            # Save to temporary file for review (will be cleaned up)
            import tempfile
            temp_fd, temp_path = tempfile.mkstemp(suffix=extension)
            os.close(temp_fd)
            temp_path = Path(temp_path)
            temp_files.append(temp_path)
            
            with open(temp_path, "wb") as f:
                f.write(image_data)
            print(f"[OK] Generated image (iteration {i})")
            
            print(f"Reviewing image with {self.review_model}...")
            review = self.review_image(
                str(temp_path), user_prompt, i, visual_only, iterations
            )
            if review.score is not None:
                print(f"[OK] Score: {review.score}/10 (threshold: {self.QUALITY_THRESHOLD}/10)")
            else:
                print(f"[WARN] Review unavailable - image kept, quality not verified")
                print(f"  Reason: {review.error}")

            results["iterations"].append({
                "iteration": i,
                "critique": review.critique,
                "score": review.score,
                "reviewed": review.reviewed,
                "review_error": review.error,
                "needs_improvement": review.needs_improvement,
                "success": True
            })

            # Keep this draft even when the next refinement request fails.
            final_image_data = image_data
            results["final_score"] = review.score
            results["final_reviewed"] = review.reviewed and review.score is not None
            results["success"] = True
            results["quality_met"] = (
                results["final_reviewed"] and review.score >= self.QUALITY_THRESHOLD
                and not review.needs_improvement
            )

            if not review.needs_improvement:
                results["termination_reason"] = (
                    "quality_threshold_met" if results["quality_met"] else "review_unavailable"
                )
                if review.score is not None:
                    print(f"\n[OK] Quality meets threshold ({review.score} >= {self.QUALITY_THRESHOLD})")
                else:
                    # Regenerating cannot fix a reviewer that did not answer.
                    print(f"\n[WARN] Stopping without a verified score - review the image yourself")
                final_image_data = image_data
                results["final_score"] = review.score
                results["final_reviewed"] = review.reviewed and review.score is not None
                results["success"] = True
                results["early_stop"] = True
                break

            if i == iterations:
                results["termination_reason"] = "max_iterations"
                print(f"\n[WARN] Maximum iterations reached")
                final_image_data = image_data
                results["final_score"] = review.score
                results["final_reviewed"] = review.reviewed and review.score is not None
                results["success"] = True
                break

            print(f"\n[WARN] Review requests improvement (score: {review.score}; threshold: {self.QUALITY_THRESHOLD})")
            print(f"Improving prompt...")
            current_prompt = self.improve_prompt(user_prompt, review.critique, i + 1, visual_only)
        
        # Clean up temporary files
        for temp_file in temp_files:
            try:
                if temp_file.exists():
                    temp_file.unlink()
            except Exception:
                pass
        
        # Save only the final image to output path
        if results["success"] and final_image_data:
            with open(output_path, "wb") as f:
                f.write(final_image_data)
            results["final_image"] = str(output_path)
            print(f"\n[OK] Final image: {output_path}")
        
        log_path = output_dir / f"{base_name}_review_log.json"
        log_path.write_text(json.dumps(results, indent=2), encoding="utf-8")
        print(f"[OK] Review log: {log_path}")

        print(f"\n{'='*60}")
        print(f"Generation Complete!")
        if results["final_score"] is not None:
            print(f"Final Score: {results['final_score']}/10")
        else:
            print(f"Final Score: unavailable (the review did not produce one)")
        if results["early_stop"]:
            success_count = len([r for r in results['iterations'] if r.get('success')])
            print(f"Iterations Used: {success_count}/{iterations} (early stop)")
        print(f"{'='*60}\n")
        
        return results


def main():
    """Command-line interface."""
    parser = argparse.ArgumentParser(
        description="Generate presentation slides or visuals using Nano Banana 2 AI",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  # Generate a full slide (for PDF workflow)
  python generate_slide_image_ai.py "Title: Machine Learning Basics\\nKey points: supervised learning, neural networks, deep learning" -o slide_01.png
  
  # Generate just a visual/figure (for PPT workflow)
  python generate_slide_image_ai.py "Neural network architecture diagram with input, hidden, and output layers" -o figure.png --visual-only
  
  # With reference images attached (Nano Banana 2 will see these)
  python generate_slide_image_ai.py "Create a slide explaining this chart with key insights" -o slide.png --attach chart.png
  python generate_slide_image_ai.py "Combine these images into a comparison slide" -o compare.png --attach before.png --attach after.png
  
  # With custom iterations
  python generate_slide_image_ai.py "Title slide for AI Conference 2025" -o title.png --iterations 2
  
  # Verbose output
  python generate_slide_image_ai.py "Data flow diagram" -o flow.png -v

Environment:
  OPENROUTER_API_KEY    OpenRouter API key (required)
        """
    )
    
    parser.add_argument("prompt", help="Description of the slide or visual to generate")
    parser.add_argument("-o", "--output", required=True, help="Output image path")
    parser.add_argument("--attach", action="append", dest="attachments", metavar="IMAGE",
                       help="Attach image file(s) as context for generation (can use multiple times)")
    parser.add_argument("--visual-only", action="store_true",
                       help="Generate just the visual/figure (for PPT workflow)")
    parser.add_argument("--iterations", type=int, default=2,
                       help="Maximum refinement iterations (default: 2)")
    parser.add_argument("--api-key", help="OpenRouter API key (or set OPENROUTER_API_KEY)")
    parser.add_argument("-v", "--verbose", action="store_true", help="Verbose output")
    
    args = parser.parse_args()
    
    # Resolves --api-key, the environment, then any .env file
    api_key = _resolve_api_key(args.api_key)
    if not api_key:
        print("Error: OPENROUTER_API_KEY not found")
        print("\nSet it with:")
        print("  export OPENROUTER_API_KEY='your_api_key'")
        print("\nOr add OPENROUTER_API_KEY=your_api_key to a .env file")
        print("Or provide via --api-key flag")
        sys.exit(1)
    
    if args.iterations < 1 or args.iterations > 2:
        print("Error: Iterations must be between 1 and 2")
        sys.exit(1)
    
    # Validate attachments exist
    if args.attachments:
        for att in args.attachments:
            if not Path(att).exists():
                print(f"Error: Attachment file not found: {att}")
                sys.exit(1)
    
    try:
        generator = SlideImageGenerator(api_key=api_key, verbose=args.verbose)
        results = generator.generate_slide(
            user_prompt=args.prompt,
            output_path=args.output,
            visual_only=args.visual_only,
            iterations=args.iterations,
            attachments=args.attachments
        )
        
        if results["success"]:
            print(f"\n[OK] Success! Image saved to: {args.output}")
            if not results["quality_met"]:
                print("[WARN] Quality criteria not verified or not met; inspect the saved draft.")
            sys.exit(0)
        else:
            print(f"\n[FAIL] Generation failed. Check review log for details.")
            sys.exit(1)
    except Exception as e:
        print(f"\n[FAIL] Error: {str(e)}")
        sys.exit(1)


if __name__ == "__main__":
    main()
canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill.Stars: 198kUpdated Claude CodeCodexVerifiedInstall