app.py

import os
import torch
import gradio as gr
import spaces
import random
import numpy as np
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download

from diffusers.utils import logging
from PIL import Image

from ovis_image.model.tokenizer import build_ovis_tokenizer
from ovis_image.model.autoencoder import load_ae
from ovis_image.model.hf_embedder import OvisEmbedder
from ovis_image.model.model import OvisImageModel
from ovis_image.sampling import generate_image
from ovis_image import ovis_image_configs

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")

print("init ovis_image")
model_config = ovis_image_configs["ovis-image-7b"]
ovis_image = OvisImageModel(model_config)
ovis_image_path = hf_hub_download(
    repo_id="AIDC-AI/Ovis-Image-7B", 
    filename="ovis_image.safetensors", 
    token=hf_token,
)
model_state_dict = load_file(ovis_image_path)
missing_keys, unexpected_keys = ovis_image.load_state_dict(model_state_dict)
print(f"Load Missing Keys {missing_keys}")
print(f"Load Unexpected Keys {unexpected_keys}")
ovis_image = ovis_image.to(device=device, dtype=_dtype)
ovis_image.eval()

print("init vae")
vae_path = hf_hub_download(
    repo_id="AIDC-AI/Ovis-Image-7B", 
    filename="ae.safetensors", 
    token=hf_token,
)
autoencoder = load_ae(
    vae_path,
    model_config.autoencoder_params,
    device=device,
    dtype=_dtype,
    random_init=False,
)
autoencoder.eval()

print("init ovis")
# ovis_path = hf_hub_download(
#     repo_id="AIDC-AI/Ovis-Image-7B", 
#     subfolder="Ovis2.5-2B", 
#     token=hf_token,
# )
ovis_tokenizer = build_ovis_tokenizer(
    "AIDC-AI/Ovis2.5-2B",
)
ovis_encoder = OvisEmbedder(
    model_path="AIDC-AI/Ovis2.5-2B",
    random_init=False,
    low_cpu_mem_usage=True,
    torch_dtype=torch.bfloat16,
).to(device=device, dtype=_dtype)


@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}')
    image = generate_image(
        device=next(ovis_image.parameters()).device,
        dtype=_dtype,
        model=ovis_image,
        prompt=prompt,
        autoencoder=autoencoder,
        ovis_tokenizer=ovis_tokenizer,
        ovis_encoder=ovis_encoder,
        img_height=img_height,
        img_width=img_width,
        denoising_steps=steps,
        cfg_scale=guidance_scale,
        seed=seed,
    )
    # bring into PIL format and save
    image = image.clamp(-1, 1)
    image = image.cpu().permute(0, 2, 3, 1).float().numpy()
    image = (image * 255).round().astype("uint8")
    
    return image[0]

examples = [
    "Solar punk vehicle in a bustling city",
    "An anthropomorphic cat riding a Harley Davidson in Arizona with sunglasses and a leather jacket",
    "An elderly woman poses for a high fashion photoshoot in colorful, patterned clothes with a cyberpunk 2077 vibe",
]

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
[[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()