InstantIR / app.py
JOY-Huang's picture
Update space
9f954a0
raw
history blame
11 kB
import os
print(os.listdir('.'))
import torch
import random
import numpy as np
import gradio as gr
from PIL import Image
from torchvision import transforms
from diffusers import (
DDPMScheduler,
StableDiffusionXLPipeline
)
from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler
from diffusers.utils import convert_unet_state_dict_to_peft
from peft import LoraConfig, set_peft_model_state_dict
from transformers import (
AutoImageProcessor, AutoModel
)
from module.ip_adapter.utils import init_ip_adapter_in_unet
from module.ip_adapter.resampler import Resampler
from module.aggregator import Aggregator
from pipelines.sdxl_instantir import InstantIRPipeline, LCM_LORA_MODULES, PREVIEWER_LORA_MODULES
transform = transforms.Compose([
transforms.Resize(1024, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(1024),
])
device = "cuda" if torch.cuda.is_available() else "cpu"
sdxl_repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
instantir_repo_id = "instantx/instantir"
dinov2_repo_id = "facebook/dinov2-large"
if torch.cuda.is_available():
torch_dtype = torch.float16
else:
torch_dtype = torch.float32
print("Loading vision encoder...")
image_encoder = AutoModel.from_pretrained(dinov2_repo_id, torch_dtype=torch_dtype)
image_processor = AutoImageProcessor.from_pretrained(dinov2_repo_id)
print("Loading SDXL...")
pipe = StableDiffusionXLPipeline.from_pretrained(
sdxl_repo_id,
torch_dtype=torch.float16,
)
unet = pipe.unet
print("Initializing Aggregator...")
aggregator = Aggregator.from_unet(unet, load_weights_from_unet=False)
print("Loading LQ-Adapter...")
image_proj_model = Resampler(
dim=1280,
depth=4,
dim_head=64,
heads=20,
num_queries=64,
embedding_dim=image_encoder.config.hidden_size,
output_dim=unet.config.cross_attention_dim,
ff_mult=4
)
init_ip_adapter_in_unet(
unet,
image_proj_model,
"InstantX/InstantIR/adapter.pt",
adapter_tokens=64,
)
print("Initializing InstantIR...")
pipe = InstantIRPipeline(
pipe.vae, pipe.text_encoder, pipe.text_encoder_2, pipe.tokenizer, pipe.tokenizer_2,
unet, aggregator, pipe.scheduler, feature_extractor=image_processor, image_encoder=image_encoder,
)
# Add Previewer LoRA.
lora_state_dict, alpha_dict = StableDiffusionXLPipeline.lora_state_dict(
"InstantX/InstantIR/previewer_lora_weights.bin",
# weight_name="previewer_lora_weights.bin",
)
unet_state_dict = {
f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
}
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
lora_state_dict = dict()
for k, v in unet_state_dict.items():
if "ip" in k:
k = k.replace("attn2", "attn2.processor")
lora_state_dict[k] = v
else:
lora_state_dict[k] = v
if alpha_dict:
lora_alpha = next(iter(alpha_dict.values()))
else:
lora_alpha = 1
print(f"use lora alpha {lora_alpha}")
lora_config = LoraConfig(
r=64,
target_modules=PREVIEWER_LORA_MODULES,
lora_alpha=lora_alpha,
lora_dropout=0.0,
)
# Add LCM LoRA.
lora_state_dict, alpha_dict = StableDiffusionXLPipeline.lora_state_dict(
"latent-consistency/lcm-lora-sdxl"
)
unet_state_dict = {
f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
}
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
if alpha_dict:
lora_alpha = next(iter(alpha_dict.values()))
else:
lora_alpha = 1
print(f"use lora alpha {lora_alpha}")
lora_config = LoraConfig(
r=64,
target_modules=LCM_LORA_MODULES,
lora_alpha=lora_alpha,
lora_dropout=0.0,
)
unet.add_adapter(lora_config, "lcm")
incompatible_keys = set_peft_model_state_dict(unet, unet_state_dict, adapter_name="lcm")
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
missing_keys = getattr(incompatible_keys, "missing_keys", None)
if unexpected_keys:
raise ValueError(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
unet.disable_adapters()
pipe.scheduler = DDPMScheduler.from_pretrained(
sdxl_repo_id,
subfolder="scheduler"
)
lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)
# Load weights.
print("Loading checkpoint...")
aggregator_state_dict = torch.load(
"InstantX/InstantIR/aggregator.pt",
map_location="cpu"
)
aggregator.load_state_dict(aggregator_state_dict, strict=True)
aggregator.to(dtype=torch.float16)
unet.to(dtype=torch.float16)
pipe=pipe.to(device)
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
def unpack_pipe_out(preview_row, index):
return preview_row[index][0]
def dynamic_preview_slider(sampling_steps):
print(sampling_steps)
return gr.Slider(label="Restoration Previews", value=sampling_steps-1, minimum=0, maximum=sampling_steps-1, step=1)
def dynamic_guidance_slider(sampling_steps):
return gr.Slider(label="Start Free Rendering", value=sampling_steps, minimum=0, maximum=sampling_steps, step=1)
def show_final_preview(preview_row):
return preview_row[-1][0]
# @spaces.GPU #[uncomment to use ZeroGPU]
def instantir_restore(lq, prompt="", steps=30, cfg_scale=7.0, guidance_end=1.0, creative_restoration=False, seed=3407):
if creative_restoration:
if "lcm" not in pipe.unet.active_adapters():
pipe.unet.set_adapter('lcm')
else:
if "previewer" not in pipe.unet.active_adapters():
pipe.unet.set_adapter('previewer')
if isinstance(guidance_end, int):
guidance_end = guidance_end / steps
with torch.no_grad(): lq = [transform(lq)]
generator = torch.Generator(device=device).manual_seed(seed)
out = pipe(
prompt=[prompt]*len(lq),
image=lq,
ip_adapter_image=[lq],
num_inference_steps=steps,
generator=generator,
controlnet_conditioning_scale=1.0,
# negative_original_size=(256,256),
# negative_target_size=(1024,1024),
negative_prompt=[""]*len(lq),
guidance_scale=cfg_scale,
control_guidance_end=guidance_end,
# control_guidance_start=0.5,
previewer_scheduler=lcm_scheduler,
return_dict=False,
save_preview_row=True,
# reference_latent = reference_latents,
# output_type='pt'
)
for i, preview_img in enumerate(out[1]):
preview_img.append(f"preview_{i}")
return out[0][0], out[1]
examples = [
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
"An astronaut riding a green horse",
"A delicious ceviche cheesecake slice",
]
css="""
#col-container {
margin: 0 auto;
max-width: 640px;
}
"""
with gr.Blocks(css=css) as demo:
gr.Markdown(
"""
# InstantIR: Blind Image Restoration with Instant Generative Reference.
### **Official 🤗 Gradio demo of [InstantIR](https://arxiv.org/abs/2410.06551).**
### **InstantIR can not only help you restore your broken image, but also capable of imaginative re-creation following your text prompts. See advance usage for more details!**
## Basic usage: revitalize your image
1. Upload an image you want to restore;
2. Optionally, tune the `Steps` `CFG Scale` parameters. Typically higher steps lead to better results, but less than 50 is recommended for efficiency;
3. Click `InstantIR magic!`.
""")
with gr.Row():
lq_img = gr.Image(label="Low-quality image", type="pil")
with gr.Column(elem_id="col-container"):
with gr.Row():
steps = gr.Number(label="Steps", value=20, step=1)
cfg_scale = gr.Number(label="CFG Scale", value=7.0, step=0.1)
seed = gr.Number(label="Seed", value=42, step=1)
# guidance_start = gr.Slider(label="Guidance Start", value=1.0, minimum=0.0, maximum=1.0, step=0.05)
guidance_end = gr.Slider(label="Start Free Rendering", value=20, minimum=0, maximum=20, step=1)
prompt = gr.Textbox(
label="Restoration prompts (Optional)", show_label=False,
placeholder="Restoration prompts (Optional)", value='',
# container=False,
)
mode = gr.Checkbox(label="Creative Restoration", value=False)
# with gr.Accordion("Advanced Settings", open=False):
with gr.Row():
with gr.Row():
restore_btn = gr.Button("InstantIR magic!")
clear_btn = gr.ClearButton()
index = gr.Slider(label="Restoration Previews", value=19, minimum=0, maximum=19, step=1)
with gr.Row():
output = gr.Image(label="InstantIR restored", type="pil")
preview = gr.Image(label="Preview", type="pil")
# gr.Examples(
# examples = examples,
# inputs = [prompt]
# )
# gr.on(
# triggers=[restore_btn.click, prompt.submit],
# fn = infer,
# inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
# outputs = [result, seed]
# )
pipe_out = gr.Gallery(visible=False)
clear_btn.add([lq_img, output, preview])
restore_btn.click(instantir_restore, inputs=[lq_img, prompt, steps, cfg_scale, guidance_end, mode, seed], outputs=[output, pipe_out], api_name="InstantIR")
steps.change(dynamic_guidance_slider, inputs=steps, outputs=guidance_end)
output.change(dynamic_preview_slider, inputs=steps, outputs=index)
index.release(unpack_pipe_out, inputs=[pipe_out, index], outputs=preview)
output.change(show_final_preview, inputs=pipe_out, outputs=preview)
gr.Markdown(
"""
## Advance usage:
### Browse restoration variants:
1. After InstantIR processing, drag the `Restoration Previews` slider to explore other in-progress versions;
2. If you like one of them, set the `Start Free Rendering` slider to the same value to get a more refined result.
### Creative restoration:
1. Check the `Creative Restoration` checkbox;
2. Input your text prompts in the `Restoration prompts` textbox;
3. Set `Start Free Rendering` slider to a medium value (around half of the `steps`) to provide adequate room for InstantIR creation.
## Examples
Here are some examplar usage of InstantIR:
""")
# examples = gr.Gallery(label="Examples")
gr.Markdown(
"""
## Citation
If InstantIR is helpful to your work, please cite our paper via:
```
@article{huang2024instantir,
title={InstantIR: Blind Image Restoration with Instant Generative Reference},
author={Huang, Jen-Yuan and Wang, Haofan and Wang, Qixun and Bai, Xu and Ai, Hao and Xing, Peng and Huang, Jen-Tse},
journal={arXiv preprint arXiv:2410.06551},
year={2024}
}
```
""")
demo.queue().launch(debug=True)