app_diffusers.py

import os
import torch
import gradio as gr
import spaces
import random
import numpy as np

from diffusers.utils import logging
from PIL import Image

from diffusers import OvisImagePipeline


logging.set_verbosity_error()

# DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MAX_SEED = np.iinfo(np.int32).max

device = "cuda"
_dtype = torch.bfloat16
hf_token = os.getenv("HF_TOKEN")

pipe = OvisImagePipeline.from_pretrained(
    "AIDC-AI/Ovis-Image-7B", 
    token=hf_token, 
    torch_dtype=torch.bfloat16
)
pipe.to("cuda")

@spaces.GPU(duration=75)
def generate(prompt, img_height=1024, img_width=1024, seed=42, steps=50, guidance_scale=5.0):
    print(f'inference with prompt : {prompt}, size: {img_height}x{img_width}, seed : {seed}, step : {steps}, cfg : {guidance_scale}')
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt, 
        negative_prompt="", 
        height=img_height,
        width=img_width,
        num_inference_steps=steps, 
        true_cfg_scale=guidance_scale,
        generator=generator,
    ).images[0]
    return image

examples = [
    "Child lying on bed with balloons, streamers, and plush toys, surreal dreamlike scene. Close-up, centered on the child and surrounding objects. Soft, diffused lighting.",
    "A trendy, playful 3D render of the text \"OVIS-IMAGE\" designed as glossy, inflated balloon letters. The typography is bubbly and rounded. The material is a shiny, reflective plastic in vibrant Alibaba Orange. The text is floating in a bright, clean white studio space with soft pastel lighting. There are subtle reflections of the studio softboxes on the curves of the balloons. High-fashion retail aesthetic, pop art style, C4D render, cute and energetic.",
    "一张写实风格的现代化教室场景摄影图。画面焦点集中在前方墙壁正中央的一块洁白明亮的白板上，白板上用清晰、工整的黑色马克笔手写体写着多行文字：最上方是巨大的英文标题“OVIS-IMAGE”，紧接着下方依次分行写着中文说明：“7B文生图模型”、“无需专有模型即可实现双语渲染”、“可在消费级显卡部署”。白板下方的铝合金笔槽内摆放着几支直立或横躺的彩色白板笔（蓝、红、黄、绿）。前景是两位背对镜头的学生（呈虚化状态，营造景深感），坐在木质课桌前，右侧课桌上还放着一块折叠的棕色毛巾。左侧隐约可见部分绿色黑板，顶部有一盏明亮的长条日光灯提供照明。整体光线柔和自然，构图对称，对焦精准在白板文字上，营造出真实的教育和技术演示氛围。",
    "微距摄影，一只鲜艳的红色瓢虫停留在翠绿的叶子上，特写镜头。叶子表面分布着晶莹剔透的水珠（晨露），瓢虫红色的甲壳光滑如镜，反射着柔和的自然光，能够清晰看到瓢虫腿部和触角的细微纹理。背景是梦幻的绿色虚化（散景），突显主体，极度逼真，8k分辨率，高清晰度，电影级质感。",
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 520px;
}
"""

with gr.Blocks(css=css) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""# Ovis-Image
Built upon [Ovis-U1](https://huggingface.co/spaces/AIDC-AI/Ovis-U1-3B), Ovis-Image is a 7B text-to-image model specifically optimized for high-quality text rendering under stringent computational constraints.

[[code](https://github.com/AIDC-AI/Ovis-Image)] [[model](https://huggingface.co/AIDC-AI/Ovis-Image-7B)]
        """)
        
        with gr.Row():
            
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt here",
                container=False,
            )
            
            run_button = gr.Button("Run", scale=0)
        
        result = gr.Image(label="Result", show_label=False)
        
        with gr.Accordion("Advanced Settings", open=False):
            with gr.Row():

                img_height = gr.Slider(
                    label="Image Height",
                    minimum=256,
                    maximum=2048,
                    step=32,
                    value=1024,
                )
  
                img_width = gr.Slider(
                    label="Image Width",
                    minimum=256,
                    maximum=2048,
                    step=32,
                    value=1024,
                )
                        
            with gr.Row():

                guidance_scale = gr.Slider(
                    label="Guidance Scale",
                    minimum=1,
                    maximum=14,
                    step=0.1,
                    value=5.0,
                )
  
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=100,
                    step=1,
                    value=50,
                )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=42,
            )
        
        gr.Examples(
            examples = examples,
            fn = generate,
            inputs = [prompt],
            outputs = [result],
            cache_examples="lazy"
        )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn = generate,
        inputs = [prompt, img_height, img_width, seed, num_inference_steps, guidance_scale],
        outputs = [result]
    )

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