vibe-aigc 0.6.0__py3-none-any.whl → 0.6.2__py3-none-any.whl

vibe_aigc/fidelity.py

@@ -0,0 +1,401 @@
+"""Fidelity Measurement — Creative Unit Tests for vibe-aigc.
+
+Paper Section 6: "The Verification Crisis... no universal unit test for a 'cinematic atmosphere'"
+Paper Section 7: "We need 'Creative Unit Tests'"
+
+This module measures how well vibe-aigc achieves user intent:
+1. Intent Alignment: Does output match the vibe?
+2. Consistency: Same prompt → similar results?
+3. Quality Distribution: What's the score spread?
+4. Refinement Efficacy: Does feedback improve scores?
+"""
+
+import asyncio
+import statistics
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+from datetime import datetime
+from pathlib import Path
+import json
+
+from .vibe_backend import VibeBackend, GenerationRequest, GenerationResult
+from .discovery import Capability
+from .vlm_feedback import VLMFeedback, FeedbackResult
+
+
+@dataclass
+class FidelityScore:
+    """Score for a single generation."""
+    prompt: str
+    output_url: str
+    quality_score: float
+    feedback: str
+    strengths: List[str]
+    weaknesses: List[str]
+    attempt_number: int
+    timestamp: str
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "prompt": self.prompt,
+            "output_url": self.output_url,
+            "quality_score": self.quality_score,
+            "feedback": self.feedback,
+            "strengths": self.strengths,
+            "weaknesses": self.weaknesses,
+            "attempt_number": self.attempt_number,
+            "timestamp": self.timestamp
+        }
+
+
+@dataclass
+class FidelityReport:
+    """Complete fidelity report for a prompt."""
+    prompt: str
+    capability: str
+    num_runs: int
+    scores: List[FidelityScore]
+    
+    # Statistics
+    mean_score: float = 0.0
+    std_dev: float = 0.0
+    min_score: float = 0.0
+    max_score: float = 0.0
+    
+    # Refinement analysis
+    first_attempt_mean: float = 0.0
+    refined_attempts_mean: float = 0.0
+    refinement_improvement: float = 0.0
+    
+    # Common patterns
+    common_strengths: List[str] = field(default_factory=list)
+    common_weaknesses: List[str] = field(default_factory=list)
+    
+    def compute_statistics(self) -> None:
+        """Compute statistics from scores."""
+        if not self.scores:
+            return
+        
+        quality_scores = [s.quality_score for s in self.scores]
+        
+        self.mean_score = statistics.mean(quality_scores)
+        self.std_dev = statistics.stdev(quality_scores) if len(quality_scores) > 1 else 0.0
+        self.min_score = min(quality_scores)
+        self.max_score = max(quality_scores)
+        
+        # Refinement analysis
+        first_attempts = [s.quality_score for s in self.scores if s.attempt_number == 1]
+        refined_attempts = [s.quality_score for s in self.scores if s.attempt_number > 1]
+        
+        if first_attempts:
+            self.first_attempt_mean = statistics.mean(first_attempts)
+        if refined_attempts:
+            self.refined_attempts_mean = statistics.mean(refined_attempts)
+            self.refinement_improvement = self.refined_attempts_mean - self.first_attempt_mean
+        
+        # Common patterns
+        all_strengths = []
+        all_weaknesses = []
+        for s in self.scores:
+            all_strengths.extend(s.strengths)
+            all_weaknesses.extend(s.weaknesses)
+        
+        # Count frequency
+        from collections import Counter
+        strength_counts = Counter(all_strengths)
+        weakness_counts = Counter(all_weaknesses)
+        
+        self.common_strengths = [s for s, _ in strength_counts.most_common(5)]
+        self.common_weaknesses = [w for w, _ in weakness_counts.most_common(5)]
+    
+    def summary(self) -> str:
+        """Human-readable summary."""
+        lines = [
+            "=" * 60,
+            "FIDELITY REPORT",
+            "=" * 60,
+            "",
+            f"Prompt: {self.prompt[:50]}...",
+            f"Capability: {self.capability}",
+            f"Runs: {self.num_runs}",
+            "",
+            "QUALITY SCORES:",
+            f"  Mean: {self.mean_score:.2f}/10",
+            f"  Std Dev: {self.std_dev:.2f}",
+            f"  Range: {self.min_score:.1f} - {self.max_score:.1f}",
+            "",
+            "REFINEMENT EFFICACY:",
+            f"  First attempt mean: {self.first_attempt_mean:.2f}",
+            f"  Refined attempts mean: {self.refined_attempts_mean:.2f}",
+            f"  Improvement: {self.refinement_improvement:+.2f}",
+            "",
+            "COMMON STRENGTHS:",
+        ]
+        for s in self.common_strengths[:3]:
+            lines.append(f"  + {s}")
+        
+        lines.append("")
+        lines.append("COMMON WEAKNESSES:")
+        for w in self.common_weaknesses[:3]:
+            lines.append(f"  - {w}")
+        
+        lines.append("")
+        lines.append("=" * 60)
+        
+        # Verdict
+        if self.mean_score >= 7.0:
+            lines.append("VERDICT: HIGH FIDELITY - System achieves intent well")
+        elif self.mean_score >= 5.0:
+            lines.append("VERDICT: MODERATE FIDELITY - Room for improvement")
+        else:
+            lines.append("VERDICT: LOW FIDELITY - Significant gap from intent")
+        
+        if self.refinement_improvement > 0.5:
+            lines.append(f"REFINEMENT: EFFECTIVE (+{self.refinement_improvement:.1f} improvement)")
+        elif self.refinement_improvement < -0.5:
+            lines.append(f"REFINEMENT: COUNTERPRODUCTIVE ({self.refinement_improvement:.1f})")
+        else:
+            lines.append("REFINEMENT: MARGINAL EFFECT")
+        
+        lines.append("=" * 60)
+        
+        return "\n".join(lines)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "prompt": self.prompt,
+            "capability": self.capability,
+            "num_runs": self.num_runs,
+            "scores": [s.to_dict() for s in self.scores],
+            "statistics": {
+                "mean": self.mean_score,
+                "std_dev": self.std_dev,
+                "min": self.min_score,
+                "max": self.max_score,
+            },
+            "refinement": {
+                "first_attempt_mean": self.first_attempt_mean,
+                "refined_mean": self.refined_attempts_mean,
+                "improvement": self.refinement_improvement,
+            },
+            "patterns": {
+                "common_strengths": self.common_strengths,
+                "common_weaknesses": self.common_weaknesses,
+            }
+        }
+
+
+class FidelityBenchmark:
+    """Benchmark for measuring vibe-aigc fidelity.
+    
+    Usage:
+        benchmark = FidelityBenchmark(comfyui_url="http://192.168.1.143:8188")
+        await benchmark.initialize()
+        
+        report = await benchmark.run(
+            prompt="cyberpunk samurai in neon rain",
+            capability=Capability.TEXT_TO_IMAGE,
+            num_runs=5
+        )
+        
+        print(report.summary())
+    """
+    
+    def __init__(
+        self,
+        comfyui_url: str = "http://127.0.0.1:8188",
+        max_attempts_per_run: int = 2,
+        quality_threshold: float = 7.0
+    ):
+        self.backend = VibeBackend(
+            comfyui_url=comfyui_url,
+            enable_vlm=True,
+            max_attempts=max_attempts_per_run,
+            quality_threshold=quality_threshold
+        )
+        self._initialized = False
+    
+    async def initialize(self) -> None:
+        """Initialize the benchmark."""
+        await self.backend.initialize()
+        self._initialized = True
+    
+    async def run(
+        self,
+        prompt: str,
+        capability: Capability = Capability.TEXT_TO_IMAGE,
+        num_runs: int = 5,
+        **kwargs
+    ) -> FidelityReport:
+        """Run the fidelity benchmark.
+        
+        Args:
+            prompt: The prompt to test
+            capability: What to generate
+            num_runs: How many times to run
+            **kwargs: Additional generation parameters
+        
+        Returns:
+            FidelityReport with scores and statistics
+        """
+        if not self._initialized:
+            await self.initialize()
+        
+        print(f"Running fidelity benchmark: {num_runs} runs")
+        print(f"Prompt: {prompt[:50]}...")
+        print()
+        
+        scores = []
+        
+        for i in range(num_runs):
+            print(f"Run {i+1}/{num_runs}...")
+            
+            request = GenerationRequest(
+                prompt=prompt,
+                capability=capability,
+                **kwargs
+            )
+            
+            result = await self.backend.generate(request)
+            
+            if result.success:
+                score = FidelityScore(
+                    prompt=prompt,
+                    output_url=result.output_url or "",
+                    quality_score=result.quality_score or 5.0,
+                    feedback=result.feedback or "",
+                    strengths=result.strengths or [],
+                    weaknesses=result.weaknesses or [],
+                    attempt_number=result.attempts,
+                    timestamp=datetime.now().isoformat()
+                )
+                scores.append(score)
+                print(f"  Score: {score.quality_score}/10 (attempt {score.attempt_number})")
+                if score.strengths:
+                    print(f"    Strengths: {', '.join(score.strengths[:2])}")
+                if score.weaknesses:
+                    print(f"    Weaknesses: {', '.join(score.weaknesses[:2])}")
+            else:
+                print(f"  Failed: {result.error}")
+        
+        # Build report
+        report = FidelityReport(
+            prompt=prompt,
+            capability=capability.value,
+            num_runs=num_runs,
+            scores=scores
+        )
+        report.compute_statistics()
+        
+        return report
+    
+    async def compare_prompts(
+        self,
+        prompts: List[str],
+        capability: Capability = Capability.TEXT_TO_IMAGE,
+        runs_per_prompt: int = 3
+    ) -> List[FidelityReport]:
+        """Compare fidelity across multiple prompts."""
+        reports = []
+        
+        for prompt in prompts:
+            report = await self.run(prompt, capability, runs_per_prompt)
+            reports.append(report)
+        
+        return reports
+    
+    async def test_refinement_efficacy(
+        self,
+        prompt: str,
+        capability: Capability = Capability.TEXT_TO_IMAGE,
+        num_runs: int = 5
+    ) -> Dict[str, Any]:
+        """Specifically test if VLM refinement improves quality.
+        
+        Runs with max_attempts=1 (no refinement) vs max_attempts=3 (with refinement)
+        """
+        print("Testing refinement efficacy...")
+        print()
+        
+        # Without refinement
+        print("Phase 1: Without refinement (max_attempts=1)")
+        self.backend.max_attempts = 1
+        no_refine_scores = []
+        
+        for i in range(num_runs):
+            result = await self.backend.generate(GenerationRequest(
+                prompt=prompt,
+                capability=capability
+            ))
+            if result.success:
+                no_refine_scores.append(result.quality_score or 5.0)
+                print(f"  Run {i+1}: {result.quality_score}/10")
+        
+        # With refinement
+        print()
+        print("Phase 2: With refinement (max_attempts=3)")
+        self.backend.max_attempts = 3
+        with_refine_scores = []
+        
+        for i in range(num_runs):
+            result = await self.backend.generate(GenerationRequest(
+                prompt=prompt,
+                capability=capability
+            ))
+            if result.success:
+                with_refine_scores.append(result.quality_score or 5.0)
+                print(f"  Run {i+1}: {result.quality_score}/10 (attempts: {result.attempts})")
+        
+        # Analysis
+        no_refine_mean = statistics.mean(no_refine_scores) if no_refine_scores else 0
+        with_refine_mean = statistics.mean(with_refine_scores) if with_refine_scores else 0
+        improvement = with_refine_mean - no_refine_mean
+        
+        return {
+            "prompt": prompt,
+            "without_refinement": {
+                "scores": no_refine_scores,
+                "mean": no_refine_mean,
+            },
+            "with_refinement": {
+                "scores": with_refine_scores,
+                "mean": with_refine_mean,
+            },
+            "improvement": improvement,
+            "refinement_effective": improvement > 0.5
+        }
+
+
+# =============================================================================
+# CONVENIENCE FUNCTIONS
+# =============================================================================
+
+async def measure_fidelity(
+    prompt: str,
+    comfyui_url: str = "http://127.0.0.1:8188",
+    num_runs: int = 5
+) -> FidelityReport:
+    """Quick fidelity measurement."""
+    benchmark = FidelityBenchmark(comfyui_url=comfyui_url)
+    await benchmark.initialize()
+    return await benchmark.run(prompt, num_runs=num_runs)
+
+
+async def run_creative_unit_test(
+    prompt: str,
+    expected_min_score: float = 6.0,
+    comfyui_url: str = "http://127.0.0.1:8188",
+    num_runs: int = 3
+) -> bool:
+    """Run a creative unit test — does the system achieve minimum quality?
+    
+    Returns True if mean score >= expected_min_score
+    """
+    report = await measure_fidelity(prompt, comfyui_url, num_runs)
+    passed = report.mean_score >= expected_min_score
+    
+    print(f"Creative Unit Test: {'PASSED' if passed else 'FAILED'}")
+    print(f"  Expected: >= {expected_min_score}")
+    print(f"  Actual: {report.mean_score:.2f}")
+    
+    return passed

vibe_aigc/vibe_backend.py

@@ -14,7 +14,7 @@ This works with ANY ComfyUI setup — no hardcoded models or patterns.
 import asyncio
 import aiohttp
 from typing import Any, Dict, List, Optional
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 
 from .discovery import (
@@ -51,6 +51,9 @@ class GenerationResult:
     output_path: Optional[str] = None
     quality_score: float = 0.0
     feedback: Optional[str] = None
+    strengths: List[str] = field(default_factory=list)
+    weaknesses: List[str] = field(default_factory=list)
+    prompt_improvements: List[str] = field(default_factory=list)
     error: Optional[str] = None
     workflow_used: Optional[str] = None
     model_used: Optional[str] = None
@@ -147,6 +150,10 @@ class VibeBackend:
             import random
             request.seed = random.randint(0, 2**32 - 1)
         
+        # Special handling for TEXT_TO_VIDEO: use I2V pipeline
+        if request.capability == Capability.TEXT_TO_VIDEO:
+            return await self._generate_video_via_i2v(request)
+        
         # Try to get workflow
         workflow = await self._get_workflow(request)
         if not workflow:
@@ -188,16 +195,25 @@ class VibeBackend:
         
         # Compose from available nodes
         print(f"Composing workflow for {request.capability.value}...")
+        
+        # Build kwargs based on capability
+        kwargs = {
+            "negative_prompt": request.negative_prompt,
+            "width": request.width,
+            "height": request.height,
+            "steps": request.steps,
+            "cfg": request.cfg,
+            "seed": request.seed
+        }
+        
+        # Add frames only for video capabilities
+        if request.capability in [Capability.TEXT_TO_VIDEO, Capability.IMAGE_TO_VIDEO]:
+            kwargs["frames"] = request.frames
+        
         return self.composer.compose_for_capability(
             capability=request.capability,
             prompt=request.prompt,
-            negative_prompt=request.negative_prompt,
-            width=request.width,
-            height=request.height,
-            frames=request.frames,
-            steps=request.steps,
-            cfg=request.cfg,
-            seed=request.seed
+            **kwargs
         )
     
     async def _execute_with_feedback(
@@ -230,30 +246,65 @@ class VibeBackend:
                 return result
             
             # VLM feedback
-            if self.vlm and self.vlm.available and result.output_path:
-                feedback = self.vlm.analyze_media(
-                    Path(result.output_path), 
-                    current_prompt
-                )
-                
-                result.quality_score = feedback.quality_score
-                result.feedback = feedback.description
-                
-                if feedback.quality_score > best_score:
-                    best_score = feedback.quality_score
-                    best_result = result
+            if self.vlm and self.vlm.available and result.output_url:
+                # Download image for VLM analysis
+                feedback = None
+                temp_path = None
+                try:
+                    import tempfile
+                    import os
+                    async with aiohttp.ClientSession() as session:
+                        async with session.get(result.output_url) as resp:
+                            if resp.status == 200:
+                                content = await resp.read()
+                                # Save to temp file (won't auto-delete)
+                                suffix = '.png' if 'png' in result.output_url else '.webp'
+                                fd, temp_path = tempfile.mkstemp(suffix=suffix)
+                                os.write(fd, content)
+                                os.close(fd)
+                                
+                                feedback = self.vlm.analyze_media(
+                                    Path(temp_path), 
+                                    current_prompt
+                                )
+                except Exception as e:
+                    print(f"VLM feedback failed: {e}")
+                    feedback = None
+                finally:
+                    # Clean up temp file (ignore errors on Windows)
+                    if temp_path:
+                        try:
+                            import os
+                            os.unlink(temp_path)
+                        except:
+                            pass  # Windows file locking, will be cleaned up by OS
                 
-                if feedback.quality_score >= self.quality_threshold:
-                    print(f"Quality threshold met: {feedback.quality_score}/10")
+                if feedback:
+                    result.quality_score = feedback.quality_score
+                    result.feedback = feedback.description
+                    result.strengths = feedback.strengths
+                    result.weaknesses = feedback.weaknesses
+                    result.prompt_improvements = feedback.prompt_improvements
+                    
+                    if feedback.quality_score > best_score:
+                        best_score = feedback.quality_score
+                        best_result = result
+                    
+                    if feedback.quality_score >= self.quality_threshold:
+                        print(f"Quality threshold met: {feedback.quality_score}/10")
+                        result.attempts = attempt + 1
+                        return result
+                    
+                    # Refine prompt for next attempt
+                    if attempt < self.max_attempts - 1:
+                        current_prompt = self.vlm.suggest_improvements(feedback, current_prompt)
+                        print(f"Refined prompt: {current_prompt[:50]}...")
+                else:
+                    # VLM failed, return successful result
                     result.attempts = attempt + 1
                     return result
-                
-                # Refine prompt for next attempt
-                if attempt < self.max_attempts - 1:
-                    current_prompt = self.vlm.suggest_improvements(feedback, current_prompt)
-                    print(f"Refined prompt: {current_prompt[:50]}...")
             else:
-                # No VLM, return first successful result
+                # No VLM configured, return first successful result
                 result.attempts = attempt + 1
                 return result
         
@@ -341,6 +392,199 @@ class VibeBackend:
         except Exception as e:
             return GenerationResult(success=False, error=str(e))
     
+    async def _generate_video_via_i2v(self, request: GenerationRequest) -> GenerationResult:
+        """Generate video via Image-to-Video pipeline.
+        
+        Two-step process:
+        1. Generate base image with TEXT_TO_IMAGE
+        2. Animate with IMAGE_TO_VIDEO
+        """
+        print("\n[1/2] Generating base image...")
+        
+        # Step 1: Generate image
+        image_workflow = self._create_flux_image_workflow(
+            prompt=request.prompt,
+            negative=request.negative_prompt,
+            width=request.width,
+            height=request.height,
+            seed=request.seed
+        )
+        
+        image_result = await self._execute_workflow(image_workflow)
+        if not image_result.success:
+            return GenerationResult(
+                success=False,
+                error=f"Image generation failed: {image_result.error}"
+            )
+        
+        print(f"    Base image: {image_result.output_path}")
+        
+        # Step 2: Upload image and animate
+        print("\n[2/2] Animating with I2V...")
+        
+        # Download image
+        async with aiohttp.ClientSession() as session:
+            async with session.get(image_result.output_url) as resp:
+                image_data = await resp.read()
+            
+            # Upload to ComfyUI
+            form = aiohttp.FormData()
+            form.add_field('image', image_data, filename='input.png', content_type='image/png')
+            
+            async with session.post(f"{self.url}/upload/image", data=form) as resp:
+                upload_result = await resp.json()
+                uploaded_name = upload_result.get("name", "input.png")
+                print(f"    Uploaded: {uploaded_name}")
+        
+        # Create I2V workflow
+        i2v_workflow = self._create_wan_i2v_workflow(
+            uploaded_image=uploaded_name,
+            prompt=request.prompt,
+            negative=request.negative_prompt,
+            width=request.width,
+            height=request.height,
+            frames=request.frames,
+            seed=request.seed
+        )
+        
+        video_result = await self._execute_workflow(i2v_workflow)
+        if not video_result.success:
+            return GenerationResult(
+                success=False,
+                error=f"Animation failed: {video_result.error}"
+            )
+        
+        print(f"    Video: {video_result.output_path}")
+        return video_result
+    
+    def _create_flux_image_workflow(
+        self, prompt: str, negative: str, width: int, height: int, seed: int
+    ) -> Dict[str, Any]:
+        """Create FLUX image generation workflow."""
+        return {
+            "1": {
+                "class_type": "CheckpointLoaderSimple",
+                "inputs": {"ckpt_name": "flux1-dev-fp8.safetensors"}
+            },
+            "2": {
+                "class_type": "CLIPTextEncode",
+                "inputs": {"text": prompt, "clip": ["1", 1]}
+            },
+            "3": {
+                "class_type": "CLIPTextEncode",
+                "inputs": {"text": negative or "blurry, distorted, ugly", "clip": ["1", 1]}
+            },
+            "4": {
+                "class_type": "EmptyLatentImage",
+                "inputs": {"width": width, "height": height, "batch_size": 1}
+            },
+            "5": {
+                "class_type": "KSampler",
+                "inputs": {
+                    "seed": seed,
+                    "steps": 20,
+                    "cfg": 3.5,
+                    "sampler_name": "euler",
+                    "scheduler": "simple",
+                    "denoise": 1.0,
+                    "model": ["1", 0],
+                    "positive": ["2", 0],
+                    "negative": ["3", 0],
+                    "latent_image": ["4", 0]
+                }
+            },
+            "6": {
+                "class_type": "VAEDecode",
+                "inputs": {"samples": ["5", 0], "vae": ["1", 2]}
+            },
+            "7": {
+                "class_type": "SaveImage",
+                "inputs": {"images": ["6", 0], "filename_prefix": "vibe_base"}
+            }
+        }
+    
+    def _create_wan_i2v_workflow(
+        self, uploaded_image: str, prompt: str, negative: str,
+        width: int, height: int, frames: int, seed: int
+    ) -> Dict[str, Any]:
+        """Create Wan 2.1 I2V workflow."""
+        return {
+            "1": {
+                "class_type": "UNETLoader",
+                "inputs": {
+                    "unet_name": "I2V/Wan2_1-I2V-14B-480p_fp8_e4m3fn_scaled_KJ.safetensors",
+                    "weight_dtype": "fp8_e4m3fn"
+                }
+            },
+            "2": {
+                "class_type": "CLIPLoader",
+                "inputs": {
+                    "clip_name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors",
+                    "type": "wan"
+                }
+            },
+            "3": {
+                "class_type": "VAELoader",
+                "inputs": {"vae_name": "wan_2.1_vae.safetensors"}
+            },
+            "4": {
+                "class_type": "LoadImage",
+                "inputs": {"image": uploaded_image}
+            },
+            "5": {
+                "class_type": "CLIPTextEncode",
+                "inputs": {"text": prompt + ", smooth motion, cinematic", "clip": ["2", 0]}
+            },
+            "6": {
+                "class_type": "CLIPTextEncode",
+                "inputs": {"text": negative or "static, frozen, blurry, distorted", "clip": ["2", 0]}
+            },
+            "7": {
+                "class_type": "WanImageToVideo",
+                "inputs": {
+                    "positive": ["5", 0],
+                    "negative": ["6", 0],
+                    "vae": ["3", 0],
+                    "width": width,
+                    "height": height,
+                    "length": frames,
+                    "batch_size": 1,
+                    "start_image": ["4", 0]
+                }
+            },
+            "8": {
+                "class_type": "KSampler",
+                "inputs": {
+                    "seed": seed,
+                    "steps": 30,
+                    "cfg": 5.0,
+                    "sampler_name": "euler",
+                    "scheduler": "normal",
+                    "denoise": 1.0,
+                    "model": ["1", 0],
+                    "positive": ["7", 0],
+                    "negative": ["7", 1],
+                    "latent_image": ["7", 2]
+                }
+            },
+            "9": {
+                "class_type": "VAEDecode",
+                "inputs": {"samples": ["8", 0], "vae": ["3", 0]}
+            },
+            "10": {
+                "class_type": "VHS_VideoCombine",
+                "inputs": {
+                    "images": ["9", 0],
+                    "frame_rate": 16,
+                    "loop_count": 0,
+                    "filename_prefix": "vibe_i2v",
+                    "format": "image/webp",
+                    "pingpong": False,
+                    "save_output": True
+                }
+            }
+        }
+    
     def status(self) -> str:
         """Get backend status."""
         if not self._initialized:

vibe_aigc/vlm_feedback.py

@@ -118,25 +118,36 @@ class VLMFeedback:
         
         img = Image.open(image_path)
         
-        prompt = f"""You are an AI art director analyzing generated images.
+        prompt = f"""You are an expert AI art director analyzing AI-generated images for quality.
 
-Context: {context}
+Original prompt: {context}
+
+IMPORTANT: You MUST provide specific, actionable prompt improvements.
 
-Analyze this image and respond in JSON format:
+Analyze this image and respond ONLY with valid JSON (no markdown):
 {{
-    "quality_score": <1-10>,
-    "description": "<what you see>",
-    "strengths": ["<strength1>", "<strength2>"],
-    "weaknesses": ["<weakness1>", "<weakness2>"],
-    "prompt_improvements": ["<specific prompt addition>", ...],
+    "quality_score": <1-10 based on: composition, detail, prompt adherence, aesthetic quality>,
+    "description": "<brief description of what you see>",
+    "strengths": ["<specific strength 1>", "<specific strength 2>"],
+    "weaknesses": ["<specific weakness 1>", "<specific weakness 2>"],
+    "prompt_improvements": [
+        "<SPECIFIC phrase to ADD to prompt to fix weakness 1>",
+        "<SPECIFIC phrase to ADD to prompt to fix weakness 2>",
+        "<SPECIFIC quality modifier to add>"
+    ],
     "parameter_changes": {{
-        "cfg": <suggested cfg or null>,
-        "steps": <suggested steps or null>,
-        "sampler": "<suggested sampler or null>"
+        "cfg": <suggest higher/lower cfg if needed, or null>,
+        "steps": <suggest more/fewer steps if needed, or null>
     }}
 }}
 
-Be specific about what to ADD to the prompt to fix issues."""
+REQUIRED: prompt_improvements must have at least 2 specific suggestions like:
+- "add sharp focus" if blurry
+- "add dramatic shadows" if flat lighting
+- "add intricate details" if lacking detail
+- "add correct anatomy" if distorted
+
+Score guide: 1-3 poor, 4-5 mediocre, 6-7 good, 8-9 excellent, 10 perfect."""
 
         try:
             response = self.vlm.generate_content([prompt, img])

{vibe_aigc-0.6.0.dist-info → vibe_aigc-0.6.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vibe-aigc
-Version: 0.6.0
+Version: 0.6.2
 Summary: A New Paradigm for Content Generation via Agentic Orchestration
 Author: Vibe AIGC Contributors
 License-Expression: MIT

{vibe_aigc-0.6.0.dist-info → vibe_aigc-0.6.2.dist-info}/RECORD

@@ -8,6 +8,7 @@ vibe_aigc/comfyui.py,sha256=Do7kOaeiuJ0SVYsib2I5prOn5O0M_u0gPbK32cFHYg4,15663
 vibe_aigc/composer_general.py,sha256=r2aqVyjqimTnbBODnKxJs4fQBOp6JEHegqnvxeel9Zk,16358
 vibe_aigc/discovery.py,sha256=PrrSy25jLz44O_mq1BgnAgR2AkZHwjJxSqn_ODZm_do,14660
 vibe_aigc/executor.py,sha256=AWTqa7hoDljhXjp6gCekzbTGROtD8kptb0IU3mi03lw,38977
+vibe_aigc/fidelity.py,sha256=z3MWAvNa554LyaK8JF6Rb-xiH1o9_NW_nA3CSoMHB8s,13542
 vibe_aigc/knowledge.py,sha256=c1gC44rbebE_FjvKyLqPbJ-_9iJQlHoXr6yBcriZTe0,40728
 vibe_aigc/llm.py,sha256=MoFrSNnHT90hZEq_pXh1rEtIZfs1RDDVshliK9wyMbw,7879
 vibe_aigc/model_registry.py,sha256=sgvNy7RK_DSBo9W0DznlpOS3J1DrL4_vfrLAO-tSa6Y,27689
@@ -17,19 +18,19 @@ vibe_aigc/persistence.py,sha256=inrJQjmCK4LighxQSmJorR6c7OvRzx-cmEb5HCQS9PY,1061
 vibe_aigc/planner.py,sha256=hmnASmofpahNuF9ei_0DxzHxm23vYjF67u-SB_G5EcU,33129
 vibe_aigc/tools.py,sha256=Tm_NA53yJjjvCrUuZ7YVtdLAdfUgxOLm5zZzIcJYvHI,15572
 vibe_aigc/tools_multimodal.py,sha256=asSJJqF0hrD9uNiYpuieVY-lbgEXjbK3UjT20nX2Lig,20405
-vibe_aigc/vibe_backend.py,sha256=BnGxIyNcnqngRpVw_T5qoeKnsFrsh0z20x7oP9PNwkw,14304
+vibe_aigc/vibe_backend.py,sha256=LUm9t3JeGfezJTjau9XAQeRN_DmHrPX2PCjNRhGE4lQ,23808
 vibe_aigc/video.py,sha256=0fg8RUpEsaJqDskAPiGP8yuyQDVCUvIy-uLScq_BOwg,14111
 vibe_aigc/visualization.py,sha256=jDs2f1vj4k8ZnJTA_niKLBH2NMahTgWneiADlNmW24s,7143
-vibe_aigc/vlm_feedback.py,sha256=-4_QAp1uCoHEzQKYln1Zk25V-jyr-Gb4dptdlJVUBUA,10137
+vibe_aigc/vlm_feedback.py,sha256=Da26q5qmJr-vwdsstum8CTAjbedeLWAGxZfla2BS0Ko,10781
 vibe_aigc/workflow_backend.py,sha256=kMIgZgyg7O7txniBSjRzEYN6Aal9zEjfRJAO8FnW07o,12136
 vibe_aigc/workflow_composer.py,sha256=mNdrlyhq-Fi0H02_iB1mpCwL9k71gv1ST10Ftx99vW0,22586
 vibe_aigc/workflow_executor.py,sha256=mfYLOTfPmI7Upooxy07nPmlbZ-HZAfC18IaNW80G31E,20734
 vibe_aigc/workflow_registry.py,sha256=Z6gB1cA366LXqHcfqBF1od_8ySxAOt5RpKKaaZPqqUo,22359
 vibe_aigc/workflow_strategies.py,sha256=i_qqUrn-2F6lT9dNyFdTdy0NzE8ZnRNxAMl6zrOAtD8,26148
 vibe_aigc/workflows.py,sha256=uk7RjNVow6eimEdqfVQFDtLgHSkg0LUjSoa2N7C47u0,13886
-vibe_aigc-0.6.0.dist-info/licenses/LICENSE,sha256=Ir4dCTvOsbfoiOh9vYbhIKDH59S7J6qhJYZmHHICoKY,1079
-vibe_aigc-0.6.0.dist-info/METADATA,sha256=p0HacVM5u17rSaNYh5SxilbNIeFE4VFqMU1A9m_0coA,6604
-vibe_aigc-0.6.0.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
-vibe_aigc-0.6.0.dist-info/entry_points.txt,sha256=2htp4yXJMvCAQXTB39XWWwbBPP3MYUYXsqlwMeQsd7o,49
-vibe_aigc-0.6.0.dist-info/top_level.txt,sha256=Cpjz8X0WEhnhaigqxmsZSl9VxduaDspj7WuVUGGLeao,10
-vibe_aigc-0.6.0.dist-info/RECORD,,
+vibe_aigc-0.6.2.dist-info/licenses/LICENSE,sha256=Ir4dCTvOsbfoiOh9vYbhIKDH59S7J6qhJYZmHHICoKY,1079
+vibe_aigc-0.6.2.dist-info/METADATA,sha256=Tdhh5kFrhVvnvWNe3MkcXwrTW7FL_0w4rZ15rtmar34,6604
+vibe_aigc-0.6.2.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+vibe_aigc-0.6.2.dist-info/entry_points.txt,sha256=2htp4yXJMvCAQXTB39XWWwbBPP3MYUYXsqlwMeQsd7o,49
+vibe_aigc-0.6.2.dist-info/top_level.txt,sha256=Cpjz8X0WEhnhaigqxmsZSl9VxduaDspj7WuVUGGLeao,10
+vibe_aigc-0.6.2.dist-info/RECORD,,