app.py

import gradio as gr

import torch
from PIL import Image, ImageFilter, ImageEnhance, ImageDraw
from diffusers import LCMScheduler, StableDiffusionInpaintPipeline
from transformers import SegformerFeatureExtractor, SegformerForSemanticSegmentation

from tqdm import tqdm
import numpy as np
from datetime import datetime

# ideally:
# preferred_device = "cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu")
# but segformer does not work on mps lolololol
preferred_device = "cuda" if torch.cuda.is_available() else "cpu"
preferred_dtype = torch.float16 if preferred_device == 'cuda' else torch.float32
inpaint_preferred_device = "cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu")
torch.backends.cuda.matmul.allow_tf32 = True
preferred_backend = "aot_eager" if inpaint_preferred_device == "mps" else ("torch_tensorrt" if inpaint_preferred_device == "cuda" else "inductor")

seg_model_img_size = 768
seg_model_size = 0

seg_feature_extractor = SegformerFeatureExtractor.from_pretrained(f"nvidia/segformer-b{seg_model_size}-finetuned-cityscapes-{seg_model_img_size}-{seg_model_img_size}")
seg_model = SegformerForSemanticSegmentation.from_pretrained(
    f"nvidia/segformer-b{seg_model_size}-finetuned-cityscapes-{seg_model_img_size}-{seg_model_img_size}"
).to(preferred_device).to(preferred_dtype)

inpainting_pipeline = StableDiffusionInpaintPipeline.from_pretrained(
    "SimianLuo/LCM_Dreamshaper_v7",
    torch_dtype=preferred_dtype,
    safety_checker=None,
).to(inpaint_preferred_device)

inpainting_pipeline.unet = torch.compile(inpainting_pipeline.unet, backend=preferred_backend)
inpainting_pipeline.vae = torch.compile(inpainting_pipeline.vae, backend=preferred_backend)
seg_model = torch.compile(seg_model, backend=preferred_backend)

seg_working_size = (seg_model_img_size, seg_model_img_size)
repaint_working_size = (768, 768)

default_inpainting_prompt = "award-winning photo of a leafy pedestrian mall full of people, with multiracial genderqueer joggers and bicyclists and wheelchair users talking and laughing"

seg_vocabulary = seg_model.config.label2id
print(f"vocab: {seg_vocabulary}")

ban_cars_mask = [0] * len(seg_vocabulary)
banned_classes = ["car", "road", "sidewalk", "traffic light", "traffic sign"]
for c in banned_classes:
    ban_cars_mask[seg_vocabulary[c]] = 1
ban_cars_mask = np.array(ban_cars_mask, dtype=np.uint8)


def get_seg_mask(img):
    inputs = seg_feature_extractor(images=img, return_tensors="pt").to(preferred_device)
    outputs = seg_model(**inputs)
    logits = outputs.logits[0]
    mask = Image.fromarray((ban_cars_mask[ torch.argmax(logits, dim=0).cpu().numpy() ]) * 255)
    blurred_widened_mask = ImageEnhance.Contrast(mask.filter(ImageFilter.GaussianBlur(2))).enhance(9000)
    return blurred_widened_mask


def app(img, prompt, num_inference_steps, seed):
    start_time = datetime.now().timestamp()
    old_size = Image.fromarray(img).size
    img = np.array(Image.fromarray(img).resize(seg_working_size))
    mask = get_seg_mask(img)
    # mask.save("mask.jpg")
    mask_time = datetime.now().timestamp()
    #print(prompt.__class__, img.__class__, mask.__class__, img.shape, mask.shape, mask.dtype, img.dtype)
    overlay_img = inpainting_pipeline(
        prompt=prompt,
        image=Image.fromarray(img).resize(repaint_working_size),
        mask_image=(mask).resize(repaint_working_size),
        strength=1,
        num_inference_steps=num_inference_steps,
        height=repaint_working_size[0],
        width=repaint_working_size[1],
        generator=torch.manual_seed(int(seed)),
    ).images[0]
    #overlay_img.save("overlay_raw.jpg")
    end_time = datetime.now().timestamp()
    draw = ImageDraw.Draw(overlay_img)
    # replace spaces with newlines after many words to line break prompt
    prompt = " ".join([prompt.split(" ")[i] if (i+1) % 5 else prompt.split(" ")[i] + "\n" for i in range(len(prompt.split(" ")))])

    draw.text((10, 50), f"Old size: {old_size}\nTotal duration: {int(1000 * (end_time - start_time))}ms\nSegmentation {int(1000 * (mask_time - start_time))}ms / inpainting {int(1000 * (end_time - mask_time))} \n<{prompt}>", fill=(0, 255, 0))
    #overlay_img.save("overlay_with_text.jpg")
    return overlay_img

# warmup, for compiling and then for timing

for i in range(2):
    for j in tqdm(range(3 ** i)):
        app(np.array(Image.fromarray(np.zeros((1024,1024,3), dtype=np.uint8))), default_inpainting_prompt, 4, 42).save("zeros_inpainting_oneshot.jpg")

#ideally:
#iface = gr.Interface(app, gr.Image(sources=["webcam"], streaming=True), "image", live=True)
iface = gr.Interface(app, [gr.Image(), gr.Textbox(value=default_inpainting_prompt), gr.Number(minimum=1, maximum=8, value=4), gr.Number(value=42)], "image")
iface.launch()