File size: 3,564 Bytes
7acb2a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4686696
 
 
 
 
 
 
 
 
 
 
 
 
 
7acb2a5
 
4686696
 
 
 
 
 
 
 
 
 
7acb2a5
 
 
4686696
 
 
 
 
 
 
7acb2a5
 
 
 
 
 
 
4686696
 
7acb2a5
 
 
 
 
4686696
 
 
7acb2a5
 
 
 
15bf55b
7acb2a5
 
29dbe4a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="InstantX/InstantIR", filename="models/adapter.pt", local_dir=".")
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/aggregator.pt", local_dir=".")
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/previewer_lora_weights.bin", local_dir=".")

import torch
from PIL import Image

from diffusers import DDPMScheduler
from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler

from module.ip_adapter.utils import load_adapter_to_pipe
from pipelines.sdxl_instantir import InstantIRPipeline

# prepare models under ./models
instantir_path = f'./models'

# load pretrained models
pipe = InstantIRPipeline.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16)

# load adapter
load_adapter_to_pipe(
    pipe,
    f"{instantir_path}/adapter.pt",
    image_encoder_or_path = 'facebook/dinov2-large',
)

# load previewer lora
pipe.prepare_previewers(instantir_path)
pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)

# load aggregator weights
pretrained_state_dict = torch.load(f"{instantir_path}/aggregator.pt")
pipe.aggregator.load_state_dict(pretrained_state_dict)

# send to GPU and fp16
pipe.to(device='cuda', dtype=torch.float16)
pipe.aggregator.to(device='cuda', dtype=torch.float16)

PROMPT = "Photorealistic, highly detailed, hyper detailed photo - realistic maximum detail, 32k, \
ultra HD, extreme meticulous detailing, skin pore detailing, \
hyper sharpness, perfect without deformations, \
taken using a Canon EOS R camera, Cinematic, High Contrast, Color Grading. "

NEG_PROMPT = "blurry, out of focus, unclear, depth of field, over-smooth, \
sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, \
dirty, messy, worst quality, low quality, frames, painting, illustration, drawing, art, \
watermark, signature, jpeg artifacts, deformed, lowres"

def infer(prompt, input_image, steps=30, cfg_scale=7.0, guidance_end=1.0,
    creative_restoration=False, seed=3407, height=1024, width=1024):

    
    # load a broken image
    low_quality_image = Image.open(input_image).convert("RGB")

    lq = [resize_img(low_quality_image, size=(width, height))]
    generator = torch.Generator(device=device).manual_seed(seed)
    timesteps = [
        i * (1000//steps) + pipe.scheduler.config.steps_offset for i in range(0, steps)
    ]
    timesteps = timesteps[::-1]

    prompt = PROMPT if len(prompt)==0 else prompt
    neg_prompt = NEG_PROMPT
    
    # InstantIR restoration
    image = pipe(
        prompt=[prompt]*len(lq),
        image=lq,
        num_inference_steps=steps,
        generator=generator,
        timesteps=timesteps,
        negative_prompt=[neg_prompt]*len(lq),
        guidance_scale=cfg_scale,
        previewer_scheduler=lcm_scheduler,
    ).images[0]

    return image

import gradio as gr



with gr.Blocks() as demo:
    with gr.Column():
        with gr.Row():
            with gr.Column():
                lq_img = gr.Image(label="Low-quality image", type="filepath")
                with gr.Group():
                    prompt = gr.Textbox(label="Prompt", value="")
                    
                submit_btn = gr.Button("InstantIR magic!")
            output_img = gr.Image(label="InstantIR restored")
    submit_btn.click(
        fn=infer,
        inputs=[prompt, lq_img],
        outputs=[output_img]
    )
demo.launch(show_error=True)