【通义实验室】开源【文本生成图片】大模型

文本生成图片效果

文本为一首古诗：孤帆远影碧空尽，唯见长江天际流。不同风格生成的图片

模型地址

初始化pipeline

python 复制代码

task = Tasks.text_to_image_synthesis
model_id = 'damo/multi-modal_chinese_stable_diffusion_v1.0'
pipe = pipeline(task=task, model=model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)

生成图片

python 复制代码

# 反向提示词
negative_prompt = (
        "blood, gore, violence, murder, kill, dead, corpse, "
        "horrible, frightening, scary, monster, ghost, skeleton, zombie, "
        "sex, nudity, pornography, adult, erotic, mature, "
        "drugs, alcohol, smoking, tobacco, illegal, "
        "dark, night, storm, thunder, lightning, apocalypse, disaster, "
        "gun, knife, sword, bomb, explosion, firearm, "
        "mean, angry, sadistic, hostile, aggressive, bullying, "
        "dangerous, unsafe, hazardous, poison, toxic, pollution"
    )
output = pipe(
        {
            'text': '孤帆远影碧空尽，唯见长江天际流。中国画',
            'num_inference_steps': 120,
            'guidance_scale': 11,
            'negative_prompt': negative_prompt
        }
    )
cv2.imwrite('result1.png', output['output_imgs'][0])
# 输出为opencv numpy格式，转为PIL.Image
img = output['output_imgs'][0]
img = Image.fromarray(img[:,:,::-1])
img.save('result1.png')

封装为http接口的完整代码

python 复制代码

from flask import Flask, request, send_file
import io
import torch
import cv2
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from PIL import Image

app = Flask(__name__)

# 初始化pipeline
task = Tasks.text_to_image_synthesis
model_id = 'damo/multi-modal_chinese_stable_diffusion_v1.0'
pipe = pipeline(task=task, model=model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)

@app.route('/generate', methods=['POST'])
def generate_image():
    data = request.json
    text = data.get('text', '')
    guidance_scale = data.get('guidance_scale', 9)

    if not text:
        return {'error': 'No text provided'}, 400

    negative_prompt = (
        "blood, gore, violence, murder, kill, dead, corpse, "
        "horrible, frightening, scary, monster, ghost, skeleton, zombie, "
        "sex, nudity, pornography, adult, erotic, mature, "
        "drugs, alcohol, smoking, tobacco, illegal, "
        "dark, night, storm, thunder, lightning, apocalypse, disaster, "
        "gun, knife, sword, bomb, explosion, firearm, "
        "mean, angry, sadistic, hostile, aggressive, bullying, "
        "dangerous, unsafe, hazardous, poison, toxic, pollution"
    )

    output = pipe(
        {
            'text': text,
            'num_inference_steps': 120,
            'guidance_scale': guidance_scale,
            'negative_prompt': negative_prompt
        }
    )

    img = output['output_imgs'][0]
    img = Image.fromarray(img[:, :, ::-1])  # Convert BGR to RGB

    # Save image to bytes
    img_byte_arr = io.BytesIO()
    img.save(img_byte_arr, format='PNG')
    img_byte_arr.seek(0)

    return send_file(img_byte_arr, mimetype='image/png')


if __name__ == '__main__':
    app.run(debug=False, host='0.0.0.0', port=5000)

在python环境下运行代码

第一次运行会下载大模型文件，需要等待一段时间启动成功会有如下提示

csharp 复制代码

 * Running on all addresses (0.0.0.0)
 * Running on http://127.0.0.1:5000
 * Running on http://10.10.10.132:5000

使用postman测试

源码下载地址