from __future__ import annotations MODEL_ID = "Qwen/Qwen3-0.6B" _TOKENIZER = None _MODEL = None def _runtime(): global _TOKENIZER, _MODEL if _TOKENIZER is None or _MODEL is None: from transformers import AutoModelForCausalLM, AutoTokenizer _TOKENIZER = AutoTokenizer.from_pretrained(MODEL_ID) _MODEL = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="cpu") return _TOKENIZER, _MODEL def enhance_prompt(prompt: str, *, tokenizer=None, model=None) -> str: if tokenizer is None or model is None: tokenizer, model = _runtime() messages = [ {"role": "system", "content": "Rewrite the image prompt with concrete visual detail, composition, lighting, and texture. Preserve the subject and intent. Return only the rewritten prompt."}, {"role": "user", "content": str(prompt)}, ] inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True) input_ids = inputs["input_ids"] length = len(input_ids[0]) output = model.generate(**inputs, max_new_tokens=180, do_sample=False, pad_token_id=getattr(tokenizer, "eos_token_id", None)) resolved = tokenizer.decode(output[0][length:], skip_special_tokens=True).strip() return (resolved or str(prompt).strip())[:2_000]