new image generator (drops tier designs)

app.py

@@ -253,7 +253,7 @@ class PicletGeneratorService:
     # Space endpoints
     JOY_CAPTION_SPACE = "fancyfeast/joy-caption-alpha-two"
     GPT_OSS_SPACE = "amd/gpt-oss-120b-chatbot"
-    FLUX_SPACE = "black-forest-labs/FLUX.1-schnell"
+    QWEN_IMAGE_SPACE = "multimodalart/Qwen-Image-Fast"
 
     @staticmethod
     def generate_enhanced_caption(image_path: str, hf_token: str) -> str:
@@ -304,7 +304,7 @@ class PicletGeneratorService:
             print(f"Generating text...")
             result = client.predict(
                 prompt,  # message (positional)
-                "You are a helpful assistant that creates Pokemon-style monster concepts based on real-world objects.",  # system_prompt (positional)
+                "You are a helpful assistant that creates Pokémon-style monster concepts based on real-world objects.",  # system_prompt (positional)
                 0.7,  # temperature (positional)
                 api_name="/chat"
             )
@@ -335,7 +335,7 @@ class PicletGeneratorService:
         Generate complete Piclet concept from image caption
         Returns parsed concept with object name, variation, stats, etc.
         """
-        concept_prompt = f"""You are analyzing an image to create a Pokemon-style creature. Here's the image description:
+        concept_prompt = f"""You are analyzing an image to create a Pokémon-style creature. Here's the image description:
 
 "{caption}"
 
@@ -388,7 +388,7 @@ Weight: {{e.g., "15kg" or "33 lbs"}}
 {{2-3 paragraphs describing the creature's behavior, habitat, abilities, and role in its ecosystem. What does it DO? Where does it live? How does it interact with its environment? What are its natural behaviors and powers that reflect the object's real-world function? This is the creature's background story and behavioral profile.}}
 
 # Monster Image Prompt
-{{Detailed 3-4 sentence visual description for image generation. Must include: "Inspired by [specific object name]" at the start. Describe body structure, colors, textures, materials, distinctive features, personality-driven pose/expression, dynamic action or stance, environment/background setting, and atmospheric lighting. Be specific and detailed about visual elements.}}
+{{Detailed 3-4 sentence visual description for anime-style image generation. Describe body structure, colors, textures, materials, distinctive features, personality-driven pose/expression, dynamic action or stance, environment/background setting, and atmospheric lighting. Be specific and detailed about visual elements. DO NOT mention the source object name or include phrases like "Inspired by [object]".}}
 ```
 
 CRITICAL RULES:
@@ -397,8 +397,8 @@ CRITICAL RULES:
 - Variation should be meaningful and distinctive (material, style, color, context, or model variant)
 - Physical Appearance must describe the CREATURE'S BODY with references to the specific object's visual features
 - Lore & Behavior must describe WHAT THE CREATURE DOES, not how it looks
-- Monster Image Prompt must START with "Inspired by [object name]" and be detailed (3-4 sentences)
-- Monster Image Prompt must NOT include the monster's name or style prefixes like "Anime-style" or "Pokemon-style"
+- Monster Image Prompt must be a detailed (3-4 sentences) pure visual description without mentioning the source object name
+- Monster Image Prompt must NOT include the monster's name or style prefixes like "Anime-style" or "Pokémon-style"
 - Primary Type must match the object category (machina for electronics/vehicles, structure for buildings, etc.)"""
 
         response_text = PicletGeneratorService.generate_text_with_gpt(concept_prompt, hf_token)
@@ -434,15 +434,10 @@ CRITICAL RULES:
         monster_name = extract_section(concept_text, 'Monster Name')
         primary_type = extract_section(concept_text, 'Primary Type').lower()
 
-        # Extract both appearance and lore sections, combine them
+        # Extract both appearance and lore sections separately (keep them separate!)
         physical_appearance = extract_section(concept_text, 'Physical Appearance')
         lore_behavior = extract_section(concept_text, 'Lore & Behavior')
 
-        # Combine both sections with a separator for complete description
-        description = physical_appearance
-        if lore_behavior:
-            description = f"{physical_appearance}\n\n{lore_behavior}" if physical_appearance else lore_behavior
-
         image_prompt = extract_section(concept_text, 'Monster Image Prompt')
 
         # Parse physical stats
@@ -484,7 +479,8 @@ CRITICAL RULES:
             'concept': concept_text,
             'stats': {
                 'name': monster_name or 'Unknown',
-                'description': description,
+                'physicalAppearance': physical_appearance,
+                'lore': lore_behavior,
                 'tier': tier,
                 'primaryType': primary_type or 'beast',
                 'height': height,
@@ -496,55 +492,48 @@ CRITICAL RULES:
 
     @staticmethod
     def generate_piclet_image(image_prompt: str, tier: str, hf_token: str) -> dict:
-        """Generate Piclet image using Flux"""
+        """Generate Piclet image using Qwen-Image-Fast"""
         try:
-            print(f"Connecting to Flux space...")
+            print(f"Connecting to Qwen-Image-Fast space...")
             client = Client(
-                PicletGeneratorService.FLUX_SPACE,
+                PicletGeneratorService.QWEN_IMAGE_SPACE,
                 hf_token=hf_token
             )
 
-            tier_descriptions = {
-                'low': 'simple and iconic design',
-                'medium': 'detailed and well-crafted design',
-                'high': 'highly detailed and impressive design with special effects',
-                'legendary': 'highly detailed and majestic design with dramatic lighting and aura effects'
-            }
-
-            background_styles = {
-                'low': 'simple natural environment',
-                'medium': 'detailed thematic natural environment',
-                'high': 'atmospheric thematic environment with special lighting effects',
-                'legendary': 'epic thematic environment with dramatic atmospheric lighting and mystical aura'
-            }
+            # Build enhanced prompt for Pokemon-style anime art
+            full_prompt = f"{image_prompt} Pokémon anime art style, idle pose, centered, full body visible in frame."
 
-            full_prompt = f"{image_prompt}\nNow generate a Pokémon Anime image of the monster in an idle pose with a {background_styles.get(tier, background_styles['medium'])} background that complements the creature's type and characteristics. This is a {tier} tier monster with a {tier_descriptions.get(tier, tier_descriptions['medium'])}. The monster should not be attacking or in motion. The full monster must be visible and centered in the frame."
+            print(f"Generating image with Qwen-Image-Fast...")
+            print(f"Prompt: {full_prompt[:100]}...")
 
-            print(f"Generating image with prompt: {full_prompt[:100]}...")
+            # Qwen-Image-Fast API: infer(prompt, seed, randomize_seed, aspect_ratio, guidance_scale, num_inference_steps, prompt_enhance)
             result = client.predict(
-                full_prompt,  # prompt
-                0,  # seed
-                True,  # randomize_seed
-                1024,  # width
-                1024,  # height
-                4,  # num_inference_steps
+                full_prompt,     # prompt
+                0,               # seed (will be randomized)
+                True,            # randomize_seed
+                "3:4",           # aspect_ratio (768x1024 - portrait)
+                1.0,             # guidance_scale (default)
+                8,               # num_inference_steps (default, optimized with Lightning LoRA)
+                True,            # prompt_enhance (uses LLM to enhance prompt)
                 api_name="/infer"
             )
 
-            # Extract image URL and seed
+            # Qwen returns: (PIL.Image, seed) tuple
             result_data = result.data if hasattr(result, 'data') else result
             image_data = result_data[0] if isinstance(result_data, (list, tuple)) else result_data
             seed = result_data[1] if isinstance(result_data, (list, tuple)) and len(result_data) > 1 else 0
 
-            # Handle different return formats
+            # Handle different return formats (URL or PIL Image object)
             image_url = None
             if isinstance(image_data, str):
                 image_url = image_data
             elif isinstance(image_data, dict):
                 image_url = image_data.get('url') or image_data.get('path')
+            elif hasattr(image_data, 'url'):
+                image_url = image_data.url
 
             if not image_url:
-                raise Exception("Failed to extract image URL from Flux response")
+                raise Exception("Failed to extract image URL from Qwen response")
 
             return {
                 'imageUrl': image_url,