| """ |
| PromptEngineer — enriches user prompts with heritage style keywords. |
| |
| SDXL-aware: builds long prompts (~75-150 tokens) that exploit SDXL's |
| two-encoder architecture (G+XL text encoders). Negative prompts are |
| also style-aware. |
| """ |
| from __future__ import annotations |
|
|
| import logging |
| from typing import List |
|
|
| from .base import AgentClient, AgentResponse |
| from config.styles import StyleSpec |
|
|
| log = logging.getLogger(__name__) |
|
|
| SYSTEM_PROMPT = """You are an expert prompt engineer for Stable Diffusion XL, specialized in Indian heritage art. |
| |
| Your task: take a user's plain description and enrich it with concrete visual keywords |
| for the chosen heritage style. Aim for 60-120 words. Be specific about: |
| - Composition (perspective, framing) |
| - Color palette (named pigments, not hex codes) |
| - Linework and brushwork quality |
| - Iconographic motifs authentic to the style |
| - Border / frame treatment |
| - Negative prompt should list things to avoid (modern, photographic, 3D, etc.) |
| |
| Output STRICT JSON: |
| {"prompt": "<enriched prompt 60-120 words>", |
| "negative_prompt": "<negative prompt 20-40 words>", |
| "style_summary": "<3-word summary>"}""" |
|
|
|
|
| class PromptEngineer: |
| def __init__(self, client: AgentClient | None = None) -> None: |
| self.client = client or AgentClient(temperature=0.6, max_tokens=900) |
|
|
| def enrich(self, user_prompt: str, style: StyleSpec) -> AgentResponse: |
| """Return enriched prompt + negative + summary.""" |
| if self.client.enabled: |
| user_msg = ( |
| f"User prompt: {user_prompt!r}\n" |
| f"Style: {style.id} ({style.display_name})\n" |
| f"Style region: {style.region}\n" |
| f"Authentic motifs: {', '.join(style.cultural_keywords)}\n" |
| f"Reference palette: {', '.join(style.palette)}" |
| ) |
| resp = self.client.chat(system_prompt=SYSTEM_PROMPT, user_prompt=user_msg) |
| if resp.ok: |
| resp.content = self._extract_prompt_text(resp.content, user_prompt, style) |
| return resp |
|
|
| |
| return AgentResponse( |
| content=self._heuristic_enrich(user_prompt, style), |
| model=self.client.model, |
| ok=True, |
| error="heuristic_fallback", |
| ) |
|
|
| @staticmethod |
| def _extract_prompt_text(raw: str, user_prompt: str, style: StyleSpec) -> str: |
| """Parse agent JSON; if parsing fails, return a heuristic fallback.""" |
| import json, re |
| try: |
| data = json.loads(raw) |
| return data.get("prompt", raw) |
| except Exception: |
| m = re.search(r'"prompt"\s*:\s*"([^"]+)"', raw) |
| if m: |
| return m.group(1).replace("\\n", " ").replace('\\"', '"') |
| return PromptEngineer._heuristic_enrich(user_prompt, style) |
|
|
| @staticmethod |
| def build_negative(style: StyleSpec, extra: List[str] | None = None) -> str: |
| """Compose the negative prompt for a style.""" |
| base = list(style.negative_tags) |
| |
| universal = [ |
| "low quality", "jpeg artifacts", "watermark", "signature", |
| "text", "logo", "cropped", "out of frame", "duplicate", |
| "extra limbs", "deformed", "blurry", |
| ] |
| parts = base + universal |
| if extra: |
| parts.extend(extra) |
| return ", ".join(parts) |
|
|
| @staticmethod |
| def _heuristic_enrich(user_prompt: str, style: StyleSpec) -> str: |
| tags = ", ".join(style.prompt_tags) |
| return ( |
| f"{user_prompt.strip().rstrip('.')}, {tags}, " |
| "intricate detail, traditional composition, " |
| "museum-quality heritage artwork, high resolution" |
| ) |
|
|
