---

Thanks mradermacher: For creating the GGUF versions of these models

https://huggingface.co/mradermacher/Qwen3-1.7B-Flux-Prompt-GGUF

https://huggingface.co/mradermacher/Qwen3-1.7B-Flux-Prompt-i1-GGUF

---

Qwen3-1.7B-Flux-Prompt

Turn simple words into professional Flux.1 image prompts instantly.

This is a fine-tuned version of Qwen3-1.7B-Instruct, specialized in expanding short concepts into detailed, high-quality descriptions optimized for Flux.1 image generation models.

✨ Key Features

• 🚀 Lightweight & Fast: Based on the 1.7B model, it runs extremely fast even on older GPUs or CPU.
• 🧠 "Invisible" System Prompt: The specialized system prompt is baked into the tokenizer_config.json. You don't need to type complex instructions. Just input a cat, and it outputs the full prompt automatically.
• 🎨 Non-Conversational: It doesn't chat. It doesn't say "Here is your prompt". It only outputs the raw prompt, ready for your stable diffusion pipeline.
• ✅ Validated Quality: Tested thoroughly for diversity and detail (see examples below).

---

🖼️ Examples

Input: a sexy model at home bed

Output (Generated by this model):

a sexy model at home bed, close-up shot of a woman lying on a luxurious velvet bed with soft golden lighting, elegant floral decor, minimalist modern furniture, subtle candlelight flicker, realistic textures, sensual atmosphere, high-quality photography, dramatic shadows, and a calm yet intimate mood.

---

💻 How to Use

Method 1: Python (Transformers)

Since the system prompt is integrated, the usage is extremely simple:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "aifeifei798/Qwen3-1.7B-Flux-Prompt" # Replace with your actual repo name

# Load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

def generate_flux_prompt(prompt):
    
    messages = [{"role": "user", "content": prompt}]

    # Applying the template
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        # enable_thinking=False
    )
    
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
    
    # Generation with optimized parameters for Flux prompting
    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=512,      # [Key] 512 tokens are sufficient for a detailed visual description.
        do_sample=True,          # [Required] Enable sampling to allow for creative variations.
        temperature=0.7,         # [Creativity] 0.7 offers a balance between imaginative detail and coherence.
        top_p=0.9,               # [Focus] Nucleus sampling: filters out very unlikely words.
        top_k=50,                # [Stability] Limits vocabulary to top 50 tokens to prevent hallucinations.
        repetition_penalty=1.15,  # [Variety] Slight penalty to reduce repetitive phrases without breaking grammar.
        no_repeat_ngram_size=3, 
    )
    
    # Extracting the newly generated tokens only
    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 
    
    # Parsing logic to handle potential reasoning content (if using a reasoning-capable base model)
    # Looks for the </think> token ID (usually 151668 in Qwen architecture)
    try:
        index = len(output_ids) - output_ids[::-1].index(151668)
    except ValueError:
        index = 0 # No thinking trace found, output starts from the beginning

    # Decode the final prompt
    # thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip()
    content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip()
    
    return content

# Example Usage
# if __name__ == "__main__":
    # user_input = "a sexy model at new york"
    # print(f"Input: {user_input}\n")
    # print(f"Generated Flux Prompt:\n{generate_flux_prompt(user_input)}")

---

🔧 Training Details

• Base Model: Qwen/Qwen3-1.7B-Instruct
• Dataset: flux_prompt (Alpaca format)
• Fine-tuning Framework: Unsloth
• Configuration:
  • temperature: 0.7 (Recommended)
  • repetition_penalty: 1.05 - 1.1 (Recommended)

---

💡 Tips for Users

• Keep it Simple: The model thrives on simple inputs like "a girl in rain" or "futuristic car".
• Sampling: Always use do_sample=True with temperature around 0.7 to get diverse results every time you run it.