app.py

import io
import requests
import numpy as np
import torch
from PIL import Image
from skimage.measure import block_reduce

import gradio as gr

from transformers import DetrFeatureExtractor, DetrForSegmentation, DetrConfig
from transformers.models.detr.feature_extraction_detr import rgb_to_id

from diffusers import StableDiffusionInpaintPipeline

# TODO: maybe need to port to `Blocks` system
# allegedly provides:
# Have multi-step interfaces, in which the output of one model becomes the 
# input to the next model, or have more flexible data flows in general.

# and:
# Change a component’s properties (for example, the choices in a dropdown) or its visibility based on user input
# https://huggingface.co/course/chapter9/7?fw=pt

torch.inference_mode()
torch.no_grad()

def load_segmentation_models(model_name: str = 'facebook/detr-resnet-50-panoptic'):
    feature_extractor = DetrFeatureExtractor.from_pretrained(model_name)
    model = DetrForSegmentation.from_pretrained(model_name)
    cfg = DetrConfig.from_pretrained(model_name)

    return feature_extractor, model, cfg

def load_diffusion_pipeline(model_name: str = 'runwayml/stable-diffusion-inpainting'):
    return StableDiffusionInpaintPipeline.from_pretrained(
        model_name,
        revision='fp16',
        torch_dtype=torch.float16
    )

def get_device(try_cuda=True):
    return torch.device('cuda' if try_cuda and torch.cuda.is_available() else 'cpu')

def min_pool(x: torch.Tensor, kernel_size: int):
    pad_size = (kernel_size - 1) // 2
    return -torch.nn.functional.max_pool2d(-x, kernel_size, (1, 1), padding=pad_size) 

def max_pool(x: torch.Tensor, kernel_size: int):
    pad_size = (kernel_size - 1) // 2
    return torch.nn.functional.max_pool2d(x, kernel_size, (1, 1), padding=pad_size) 

def clean_mask(mask, min_kernel: int = 5, max_kernel: int = 23):
    mask = torch.Tensor(mask[None, None]).float()
    mask = min_pool(mask, min_kernel)
    mask = max_pool(mask, max_kernel)
    mask = mask.bool().squeeze().numpy()
    return mask

device = get_device()

feature_extractor, segmentation_model, segmentation_cfg = load_segmentation_models()
# segmentation_model = segmentation_model.to(device)

pipe = load_diffusion_pipeline()
pipe = pipe.to(device)

# TODO: potentially use `gr.Gallery` to display different masks
def fn_segmentation_diffusion(prompt, mask_indices, image, max_kernel, min_kernel, num_diffusion_steps):
    mask_indices = [int(i) for i in mask_indices.split(',')]
    inputs = feature_extractor(images=image, return_tensors="pt")
    outputs = segmentation_model(**inputs)

    processed_sizes = torch.as_tensor(inputs["pixel_values"].shape[-2:]).unsqueeze(0)
    result = feature_extractor.post_process_panoptic(outputs, processed_sizes)[0]

    panoptic_seg = Image.open(io.BytesIO(result["png_string"])).resize((image.width, image.height))
    panoptic_seg = np.array(panoptic_seg, dtype=np.uint8)

    class_str = '\n'.join(segmentation_cfg.id2label[s['category_id']] for s in result['segments_info'])

    panoptic_seg_id = rgb_to_id(panoptic_seg)

    if len(mask_indices) > 0:
        mask = (panoptic_seg_id == mask_indices[0])
    for idx in mask_indices[1:]:
        mask = mask | (panoptic_seg_id == idx)
    mask = clean_mask(mask, min_kernel=min_kernel, max_kernel=max_kernel)

    masked_image = np.array(image).copy()
    masked_image[mask] = 0

    masked_image = Image.fromarray(masked_image).resize(image.size)
    mask = Image.fromarray(mask.astype(np.uint8) * 255).resize(image.size)

    if num_diffusion_steps == 0:
        return masked_image, masked_image, class_str

    STABLE_DIFFUSION_SMALL_EDGE = 512

    assert masked_image.size == mask.size
    w, h = masked_image.size
    is_width_larger = w > h
    resize_ratio = STABLE_DIFFUSION_SMALL_EDGE / (h if is_width_larger else w)

    new_width = int(w * resize_ratio) if is_width_larger else STABLE_DIFFUSION_SMALL_EDGE
    new_height = STABLE_DIFFUSION_SMALL_EDGE if is_width_larger else int(h * resize_ratio)

    new_width += 8 - (new_width % 8) if is_width_larger else 0
    new_height += 0 if is_width_larger else 8 - (new_height % 8)

    mask = mask.convert("RGB").resize((new_width, new_height))
    masked_image = masked_image.convert("RGB").resize((new_width, new_height))

    inpainted_image = pipe(
        height=new_height, 
        width=new_width, 
        prompt=prompt,
        image=masked_image, 
        mask_image=mask,
        num_inference_steps=num_diffusion_steps
    ).images[0]

    return masked_image, inpainted_image, class_str


# iface_segmentation = gr.Interface(
    # fn=fn_segmentation, 
    # inputs=[
        # "text", 
        # "text", 
        # gr.Image(value="http://images.cocodataset.org/val2017/000000039769.jpg"),
        # gr.Slider(minimum=1, maximum=99, value=23, step=2),
        # gr.Slider(minimum=1, maximum=99, value=5, step=2),
        # gr.Slider(minimum=0, maximum=100, value=50, step=1),
    # ], 
    # outputs=["text", gr.Image(type="pil"), gr.Image(type="pil"), "number", "text"]
# )

# iface_diffusion = gr.Interface(
    # fn=fn_diffusion,
    # inputs=["text", gr.Image(type='pil'), gr.Image(type='pil'), "number", "text"],
    # outputs=[gr.Image(), gr.Image(), gr.Textbox()]
# )

# iface = gr.Series(
    # iface_segmentation, iface_diffusion,
iface = gr.Interface(
    fn=fn_segmentation_diffusion,
    inputs=[
        "text",
        "text", 
        gr.Image(value="http://images.cocodataset.org/val2017/000000039769.jpg", type='pil'),
        gr.Slider(minimum=1, maximum=99, value=23, step=2),
        gr.Slider(minimum=1, maximum=99, value=5, step=2),
        gr.Slider(minimum=0, maximum=100, value=50, step=1),
    ], 
    outputs=[gr.Image(), gr.Image(), gr.Textbox(interactive=False)]
)

iface.launch()