closr

src/lib/components/MonsterGenerator/MonsterGenerator.svelte

@@ -20,24 +20,35 @@
   });
   
   // Prompt templates
-  const MONSTER_CONCEPT_PROMPT = (caption: string) => `User: Based on this description of a person/photo: "${caption}"
+  const MONSTER_CONCEPT_PROMPT = (caption: string) => `User: Based on this image caption: "${caption}"
 
-Create a unique monster concept inspired by the key visual elements. Include:
+Come up with a monster idea (in the vein of Pokémon) that incorporates properties of what the pictured object is.
+Assess how unique the pictured object is. If the image is not very unique then it should result in weak & common monster.
+Meanwhile if the image is very unique the monster should be unique and powerful.
+
+Create a creature that reflects the essence, materials, function, and characteristics of the object. Consider:
+- What the object is made of (metal, wood, fabric, etc.)
+- How it functions or is used
+- Its physical properties (hard, soft, hot, cold, etc.)
+- Its cultural or symbolic meaning
+- Its shape, texture, and visual characteristics
+
+The monster should have unique strengths and weaknesses based on these properties.
+
+Include:
 - Physical appearance and distinguishing features
 - Special abilities or powers  
 - Personality traits
 - Habitat or origin
 
-Be creative and detailed. The monster should be fantastical but somehow connected to the original image's essence.
+Assistant: **Monster Name:**`;
 
-Assistant:`;
-
-  const IMAGE_GENERATION_PROMPT = (concept: string) => `User: Convert this monster concept into a detailed image generation prompt:
+  const IMAGE_GENERATION_PROMPT = (concept: string) => `User: Convert this monster concept into a clear and succinct image generation prompt:
 "${concept}"
 
-Create a vivid, artistic prompt focusing on visual details, art style, lighting, and atmosphere. Format it as a single paragraph optimized for image generation.
+Describe the appearance of the monsters being sure to include all of its visual details, format the prompt as a single long sentence optimized for image generation.
 
-Assistant:`;
+Assistant: The`;
   
   async function handleImageSelected(file: File) {
     if (!joyCaptionClient || !rwkvClient || !fluxClient) {
@@ -113,8 +124,8 @@ Assistant:`;
     const output = await rwkvClient.predict(0, [
       conceptPrompt,
       300,   // maxTokens
-      1.2,   // temperature (more creative)
-      0.8,   // topP
+      1.0,   // temperature
+      0.3,   // topP
       0.1,   // presencePenalty
       0.1    // countPenalty
     ]);
@@ -140,10 +151,10 @@ Assistant:`;
     console.log('Generating image prompt from concept');
     
     const output = await rwkvClient.predict(0, [
-      promptGenerationPrompt,
-      150,   // maxTokens
-      0.9,   // temperature
-      0.7,   // topP
+      `${promptGenerationPrompt}\nNow generate an Anime-style image of the monster in an idle pose with a transparent background. The monster should not be attacking or in motion. The full monster must be visible within the frame.`,
+      300,   // maxTokens
+      1.0,   // temperature
+      0.3,   // topP
       0.1,   // presencePenalty
       0.1    // countPenalty
     ]);

structured_output.py

@@ -0,0 +1,36 @@
+# ruff: noqa: E501
+
+import json
+
+from pydantic import BaseModel
+
+STRUCTURED_OUTPUT_FORMAT_INSTRUCTIONS = """The output should be formatted as a JSON instance that conforms to the JSON schema below.
+
+As an example, for the schema {{"properties": {{"foo": {{"title": "Foo", "description": "a list of strings", "type": "array", "items": {{"type": "string"}}}}}}, "required": ["foo"]}}
+the object {{"foo": ["bar", "baz"]}} is a well-formatted instance of the schema. The object {{"properties": {{"foo": ["bar", "baz"]}}}} is not well-formatted.
+
+Here is the output schema:
+```
+{schema}
+```"""
+
+
+class StructuredOutputParser(BaseModel):
+    pydantic_class: type[BaseModel]
+
+    def get_schema(self, indent: int = 2) -> str:
+        # Copy schema to avoid altering original Pydantic schema.
+        schema = self.pydantic_class.model_json_schema()
+        # Iterate over fields to remove from the schema
+        for field_name in ["title", "type"]:
+            if field_name in schema:
+                schema.pop(field_name)
+        return json.dumps(schema, indent=indent, ensure_ascii=False)
+
+    def get_format_instructions(self) -> str:
+        schema_str = self.get_schema()
+        # Ensure json in context is well-formed with double quotes.
+        return STRUCTURED_OUTPUT_FORMAT_INSTRUCTIONS.format(schema=schema_str)
+
+    def parse(self, json_string: str) -> BaseModel:
+        return self.pydantic_class.model_validate_json(json_string)