app.py

from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
from diffusers import StableDiffusionInpaintPipeline as StableDiffusionInpaintPipeline
import gradio as gr
from PIL import Image
import torch
from torch import autocast
import matplotlib.pyplot as plt
import numpy as np

auth_token = os.environ.get("API_TOKEN") or True

processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")

pipe = StableDiffusionInpaintPipeline.from_pretrained(
    "runwayml/stable-diffusion-inpainting",
    revision="fp16",
    torch_dtype=torch.float16,
    use_auth_token=auth_token,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
pipe = pipe.to(device)

def pad_image(image):
    w, h = image.size
    if w == h:
        return image
    elif w > h:
        new_image = Image.new(image.mode, (w, w), (0, 0, 0))
        new_image.paste(image, (0, (w - h) // 2))
        return new_image
    else:
        new_image = Image.new(image.mode, (h, h), (0, 0, 0))
        new_image.paste(image, ((h - w) // 2, 0))
        return new_image

def process_image(image, prompt):
    inputs = processor(
        text=prompt, images=image, padding="max_length", return_tensors="pt"
    )

    # predict
    with torch.no_grad():
        outputs = model(**inputs)
        preds = outputs.logits

    pred = torch.sigmoid(preds)
    mat = pred.cpu().numpy()
    mask = Image.fromarray(np.uint8(mat * 255), "L")
    mask = mask.convert("RGB")
    mask = mask.resize(image.size)
    mask = np.array(mask)[:, :, 0]

    # normalize the mask
    mask_min = mask.min()
    mask_max = mask.max()
    mask = (mask - mask_min) / (mask_max - mask_min)
    return mask


def get_masks(prompts, img, threhsold):
    prompts = prompts.split(",")
    masks = []
    for prompt in prompts:
        mask = process_image(img, prompt)
        mask = mask > threhsold
        masks.append(mask)
    return masks


def extract_image(img, pos_prompts, neg_prompts, threshold):
    positive_masks = get_masks(pos_prompts, img, threshold)
    negative_masks = get_masks(neg_prompts, img, threshold)

    # combine masks into one masks, logic OR
    pos_mask = np.any(np.stack(positive_masks), axis=0)
    neg_mask = np.any(np.stack(negative_masks), axis=0)
    final_mask = pos_mask & ~neg_mask

    # extract the final image
    final_mask = Image.fromarray(final_mask.astype(np.uint8) * 255, "L")
    inverse_mask = np.invert(final_mask, dtype=np.uint8)
    output_image = Image.new("RGBA", img.size, (0, 0, 0, 0))
    output_image.paste(img, mask=final_mask)

    return output_image, final_mask, inverse_mask

def inpaint_image(img, mask, prompt):
    img = pad_image(img).convert("RGB").resize((512, 512))
    mask = Image.fromarray(mask*255)
    mask = pad_image(mask).convert("RGB").resize((512, 512))
    with torch.cuda.amp.autocast(True):
        inpainted_image = pipe(prompt=prompt, image=img, mask_image=mask).images[0]
    return inpainted_image
    
title = "Interactive demo: zero-shot image segmentation with CLIPSeg"
description = "Demo for using CLIPSeg, a CLIP-based model for zero- and one-shot image segmentation. To use it, simply upload an image and add a text to mask (identify in the image), or use one of the examples below and click 'submit'. Results will show up in a few seconds."
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2112.10003'>CLIPSeg: Image Segmentation Using Text and Image Prompts</a> | <a href='https://huggingface.co/docs/transformers/main/en/model_doc/clipseg'>HuggingFace docs</a></p>"

with gr.Blocks() as demo:
    gr.Markdown("# CLIPSeg: Image Segmentation Using Text and Image Prompts")
    gr.Markdown(article)
    gr.Markdown(description)

    with gr.Row():
        with gr.Column():
            input_image = gr.Image(type="pil")
            positive_prompts = gr.Textbox(
                label="Please describe what you want to identify (comma separated)"
            )
            negative_prompts = gr.Textbox(
                label="Please describe what you want to ignore (comma separated)"
            )

            input_slider_T = gr.Slider(
                minimum=0, maximum=1, value=0.4, label="Threshold"
            )
            btn_mask = gr.Button(label="Mask")

        with gr.Column():
            output_image = gr.Image(label="Result")
            output_mask = gr.Image(label="Mask")
            inverse_mask = gr.Image(label="Inverse")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(
                label="Prompt"
            )

            input_slider_S = gr.Slider(
                minimum=0, maximum=1, value=0.4, label="Image Strength"
            )
            btn_run = gr.Button(label="Run")

        with gr.Column():
            inpainted_image = gr.Image(label="Inpainted")

    btn_mask.click(
        extract_image,
        inputs=[
            input_image,
            positive_prompts,
            negative_prompts,
            input_slider_T,
        ],
        outputs=[output_image, output_mask, inverse_mask],
        api_name="mask"
    )

    btn_run.click(
        inpaint_image,
        inputs=[
            input_image,
            inverse_mask,
            prompt,
        ],
        outputs=[inpainted_image],
        api_name="run"
    )

demo.launch()