app.py

import gradio as gr
import numpy as np
import random
import torch
from PIL import Image
import os

from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256_ipadapter import StableDiffusionXLPipeline
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from kolors.models.unet_2d_condition import UNet2DConditionModel
from diffusers import AutoencoderKL, EulerDiscreteScheduler

from huggingface_hub import snapshot_download
import spaces 

device = "cuda" if torch.cuda.is_available() else "cpu"
root_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
ckpt_dir = f'{root_dir}/weights/Kolors'

snapshot_download(repo_id="Kwai-Kolors/Kolors", local_dir=ckpt_dir)
snapshot_download(repo_id="Kwai-Kolors/Kolors-IP-Adapter-Plus", local_dir=f"{root_dir}/weights/Kolors-IP-Adapter-Plus")

# Load models
text_encoder = ChatGLMModel.from_pretrained(f'{ckpt_dir}/text_encoder', torch_dtype=torch.float16).half().to(device)
tokenizer = ChatGLMTokenizer.from_pretrained(f'{ckpt_dir}/text_encoder')
vae = AutoencoderKL.from_pretrained(f"{ckpt_dir}/vae", revision=None).half().to(device)
scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler")
unet = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)

image_encoder = CLIPVisionModelWithProjection.from_pretrained(
    f'{root_dir}/weights/Kolors-IP-Adapter-Plus/image_encoder',
    ignore_mismatched_sizes=True
).to(dtype=torch.float16, device=device)

ip_img_size = 336
clip_image_processor = CLIPImageProcessor(size=ip_img_size, crop_size=ip_img_size)

pipe = StableDiffusionXLPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=scheduler,
    image_encoder=image_encoder,
    feature_extractor=clip_image_processor,
    force_zeros_for_empty_prompt=False
)

pipe = pipe.to(device)
#pipe.enable_model_cpu_offload()

if hasattr(pipe.unet, 'encoder_hid_proj'):
    pipe.unet.text_encoder_hid_proj = pipe.unet.encoder_hid_proj

pipe.load_ip_adapter(f'{root_dir}/weights/Kolors-IP-Adapter-Plus', subfolder="", weight_name=["ip_adapter_plus_general.bin"])

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024

@spaces.GPU
def infer(prompt, ip_adapter_image, negative_prompt="", seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=5.0, num_inference_steps=50, ip_adapter_scale=0.5, progress=gr.Progress(track_tqdm=True)):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
    generator = torch.Generator(device="cpu").manual_seed(seed)
    
    pipe.set_ip_adapter_scale([ip_adapter_scale])
    
    image = pipe(
        prompt=prompt,
        ip_adapter_image=[ip_adapter_image],
        negative_prompt=negative_prompt,
        height=height,
        width=width,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        num_images_per_prompt=1,
        generator=generator,
    ).images[0]
    
    return image, seed

examples = [
    ["A photo of a lavender cat", "https://upload.wikimedia.org/wikipedia/commons/thumb/4/4d/Cat_November_2010-1a.jpg/640px-Cat_November_2010-1a.jpg"],
    ["Astronaut in a jungle, cold color palette, muted colors, detailed, 8k", "https://upload.wikimedia.org/wikipedia/commons/thumb/b/b5/Astronaut_EVA.jpg/640px-Astronaut_EVA.jpg"],
    ["An astronaut riding a green horse", "https://upload.wikimedia.org/wikipedia/commons/thumb/f/f7/Haflinger_in-motion.jpg/640px-Haflinger_in-motion.jpg"],
    ["A delicious ceviche cheesecake slice", "https://upload.wikimedia.org/wikipedia/commons/thumb/9/9c/Ceviche_mixto.jpg/640px-Ceviche_mixto.jpg"],
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 720px;
}
"""
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
        # Kolors Demo
        Demo of the Kolors model with IP-Adapter integration
        """)
        
        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )
            run_button = gr.Button("Run", scale=0)
        
        with gr.Row():
            ip_adapter_image = gr.Image(label="IP-Adapter Image", type="pil")
            result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
            )
            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=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=5.0,
                )
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=100,
                    step=1,
                    value=50,
                )
            ip_adapter_scale = gr.Slider(
                label="IP-Adapter Scale",
                minimum=0.0,
                maximum=1.0,
                step=0.01,
                value=0.5,
            )
        
        gr.Examples(
            examples=examples,
            fn=infer,
            inputs=[prompt, ip_adapter_image],
            outputs=[result, seed],
            cache_examples="lazy"
        )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[prompt, ip_adapter_image, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, ip_adapter_scale],
        outputs=[result, seed]
    )

demo.queue().launch()