pipeline_sdxs.py

import gradio as gr
import numpy as np
import random
import spaces 
import torch
from diffusers import DiffusionPipeline, AutoencoderKL, UNet2DConditionModel, FlowMatchEulerDiscreteScheduler,AsymmetricAutoencoderKL
from transformers import AutoModelForCausalLM, AutoTokenizer
from typing import Optional, Union, List, Tuple
from PIL import Image

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model_repo_id = "AiArtLab/sdxs-08b"

pipe = DiffusionPipeline.from_pretrained(
    model_repo_id,
    torch_dtype=dtype,
    trust_remote_code=True
).to(device)

# НОВОЕ: Инициализация Qwen3 для рефайнинга
llm_model_id = "Qwen/Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(llm_model_id)
llm_model = AutoModelForCausalLM.from_pretrained(llm_model_id, torch_dtype="auto", device_map="auto")

MAX_SEED = np.iinfo(np.int32).max
MIN_IMAGE_SIZE = 640
MAX_IMAGE_SIZE = 1280
STEP = 64

# НОВОЕ: Настройки для LLM
END_THINK_TOKEN_ID = 151668
DEFAULT_REFINE_TEMPLATE = (
    "You are a skilled text-to-image prompt engineer whose sole function is to transform the user's input into an aesthetically optimized, detailed, and visually descriptive three-sentence output. "
    "**The primary subject (e.g., 'girl', 'dog', 'house') MUST be the main focus of the revised prompt and MUST be described in rich detail within the first sentence or two.** "
    "Output **only** the final revised prompt in **English**, with absolutely no commentary, thinking text, or surrounding quotes.\n"
    "User input prompt: {prompt}"
)

@spaces.GPU(duration=30)
def infer(
    prompt: str,
    negative_prompt: str,
    seed: int,
    randomize_seed: bool,
    width: int,
    height: int,
    guidance_scale: float,
    num_inference_steps: int,
    refine_prompt: bool,
    progress=gr.Progress(track_tqdm=True),
) -> Tuple[Image.Image, int, str]: # Возвращаем prompt в конце
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
    # НОВОЕ: Логика улучшения промпта
    if refine_prompt and prompt:
        messages = [{"role": "user", "content": DEFAULT_REFINE_TEMPLATE.format(prompt=prompt)}]
        text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
        model_inputs = tokenizer([text], return_tensors="pt").to(llm_model.device)
        
        generated_ids = llm_model.generate(**model_inputs, max_new_tokens=2048, do_sample=True, pad_token_id=tokenizer.eos_token_id)
        output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
        
        try:
            index = len(output_ids) - output_ids[::-1].index(END_THINK_TOKEN_ID)
        except ValueError:
            index = 0
        
        prompt = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n").strip()

    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        width=width,
        height=height,
        seed=seed,
    )

    image = output.images[0]
    return image, seed, prompt # Возвращаем измененный промпт

examples = [
    "A frozen river, surrounded by snow-covered trees, reflects the clear blue sky, with a warm glow from the setting sun.",
    "A young woman with striking blue eyes and pointed ears, adorned with a floral kimono and a tattoo. Her hair is styled in a braid, and she wears a pair of ears",
    "A volcano explodes, creating a skull face shadow in embers with lightning illuminating the clouds.",
    "There is a young male character standing against a vibrant, colorful graffiti wall. he is wearing a straw hat, a black jacket adorned with gold accents, and black shorts.",
    "A man with dark hair and a beard is meticulously carving an intricate design on a piece of pottery. He is wearing a traditional scarf and a white shirt, and he is focused on his work.",
    "girl, smiling, red eyes, blue hair, white shirt"
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(" # Simple Diffusion (sdxs-08b)")

        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=5,
                placeholder="Enter your prompt",
                container=False,
            )

            run_button = gr.Button("Run", scale=0, variant="primary")

        result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            # Изменено value на True
            refine_prompt = gr.Checkbox(label="Refine Prompt with Qwen3", value=True)
            
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
                value ="bad quality, low resolution"
            )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=MIN_IMAGE_SIZE,
                    maximum=MAX_IMAGE_SIZE,
                    step=STEP,
                    value=1024,
                )

                height = gr.Slider(
                    label="Height",
                    minimum=MIN_IMAGE_SIZE,
                    maximum=MAX_IMAGE_SIZE,
                    step=STEP,
                    value=MAX_IMAGE_SIZE,
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.5,
                    value=4.0,
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=40,
                )

        gr.Examples(examples=examples, inputs=[prompt])
    
    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[
            prompt,
            negative_prompt,
            seed,
            randomize_seed,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            refine_prompt,
        ],
        outputs=[result, seed, prompt], # Добавлен prompt для обновления текста в интерфейсе
    )

if __name__ == "__main__":
    demo.launch()