app.py

import os
import sys
import cv2
import numpy as np
import torch
import gradio as gr
from PIL import Image, ImageFilter, ImageDraw
from huggingface_hub import snapshot_download
from diffusers import FluxFillPipeline, FluxPriorReduxPipeline
import math
from utils.utils import get_bbox_from_mask, expand_bbox, pad_to_square, box2squre, crop_back, expand_image_mask

import os,sys
os.system("python -m pip install -e segment_anything")
os.system("python -m pip install -e GroundingDINO")
sys.path.append(os.path.join(os.getcwd(), "GroundingDINO"))
sys.path.append(os.path.join(os.getcwd(), "segment_anything"))
os.system("wget https://huggingface.co/ShilongLiu/GroundingDINO/resolve/main/groundingdino_swinb_cogcoor.pth")
os.system("wget https://huggingface.co/spaces/mrtlive/segment-anything-model/resolve/main/sam_vit_h_4b8939.pth")

import torchvision
from GroundingDINO.groundingdino.util.inference import load_model
from segment_anything import build_sam, SamPredictor 
import spaces
import GroundingDINO.groundingdino.datasets.transforms as T
from GroundingDINO.groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap



# GroundingDINO config and checkpoint
GROUNDING_DINO_CONFIG_PATH = "./GroundingDINO_SwinB.cfg.py"
GROUNDING_DINO_CHECKPOINT_PATH = "./groundingdino_swinb_cogcoor.pth"

# Segment-Anything checkpoint
SAM_ENCODER_VERSION = "vit_h"
SAM_CHECKPOINT_PATH = "./sam_vit_h_4b8939.pth"

# Building GroundingDINO inference model
groundingdino_model  = load_model(model_config_path=GROUNDING_DINO_CONFIG_PATH, model_checkpoint_path=GROUNDING_DINO_CHECKPOINT_PATH, device="cuda")
# Building SAM Model and SAM Predictor
sam = build_sam(checkpoint=SAM_CHECKPOINT_PATH)
sam.to(device="cuda")
sam_predictor = SamPredictor(sam)

def transform_image(image_pil):

    transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            T.ToTensor(),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    )
    image, _ = transform(image_pil, None)  # 3, h, w
    return image


def get_grounding_output(model, image, caption, box_threshold=0.25, text_threshold=0.25, with_logits=True):
    caption = caption.lower()
    caption = caption.strip()
    if not caption.endswith("."):
        caption = caption + "."

    with torch.no_grad():
        outputs = model(image[None], captions=[caption])
    logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
    boxes = outputs["pred_boxes"].cpu()[0]  # (nq, 4)
    logits.shape[0]

    # filter output
    logits_filt = logits.clone()
    boxes_filt = boxes.clone()
    filt_mask = logits_filt.max(dim=1)[0] > box_threshold
    logits_filt = logits_filt[filt_mask]  # num_filt, 256
    boxes_filt = boxes_filt[filt_mask]  # num_filt, 4
    logits_filt.shape[0]

    # get phrase
    tokenlizer = model.tokenizer
    tokenized = tokenlizer(caption)
    # build pred
    pred_phrases = []
    scores = []
    for logit, box in zip(logits_filt, boxes_filt):
        pred_phrase = get_phrases_from_posmap(
            logit > text_threshold, tokenized, tokenlizer)
        if with_logits:
            pred_phrases.append(
                pred_phrase + f"({str(logit.max().item())[:4]})")
        else:
            pred_phrases.append(pred_phrase)
        scores.append(logit.max().item())

    return boxes_filt, torch.Tensor(scores), pred_phrases


def get_mask(image, label):
    global groundingdino_model, sam_predictor


    image_pil = image.convert("RGB")
    transformed_image = transform_image(image_pil)


    boxes_filt, scores, pred_phrases = get_grounding_output(
        groundingdino_model, transformed_image, label
    )

    size = image_pil.size

    # process boxes
    H, W = size[1], size[0]
    for i in range(boxes_filt.size(0)):
        boxes_filt[i] = boxes_filt[i] * torch.Tensor([W, H, W, H])
        boxes_filt[i][:2] -= boxes_filt[i][2:] / 2
        boxes_filt[i][2:] += boxes_filt[i][:2]

    boxes_filt = boxes_filt.cpu()

    # nms

    nms_idx = torchvision.ops.nms(
        boxes_filt, scores, 0.8).numpy().tolist()
    boxes_filt = boxes_filt[nms_idx]
    pred_phrases = [pred_phrases[idx] for idx in nms_idx]


    image = np.array(image_pil)
    sam_predictor.set_image(image)

    transformed_boxes = sam_predictor.transform.apply_boxes_torch(
        boxes_filt, image.shape[:2]).to("cuda")

    masks, _, _ = sam_predictor.predict_torch(
        point_coords=None,
        point_labels=None,
        boxes=transformed_boxes,
        multimask_output=False,
    )
    result_mask = masks[0][0].cpu().numpy()

    result_mask = Image.fromarray(result_mask)

    return result_mask


hf_token = os.getenv("HF_TOKEN")

snapshot_download(repo_id="black-forest-labs/FLUX.1-Fill-dev", local_dir="./FLUX.1-Fill-dev", token=hf_token)
snapshot_download(repo_id="black-forest-labs/FLUX.1-Redux-dev", local_dir="./FLUX.1-Redux-dev", token=hf_token)
snapshot_download(repo_id="WensongSong/Insert-Anything", local_dir="./insertanything_model", token=hf_token)


dtype = torch.bfloat16
size = (768, 768)

pipe = FluxFillPipeline.from_pretrained(
    "./FLUX.1-Fill-dev",
    torch_dtype=dtype
).to("cuda")

pipe.load_lora_weights(
    "./insertanything_model/20250321_steps5000_pytorch_lora_weights.safetensors"
)


redux = FluxPriorReduxPipeline.from_pretrained("./FLUX.1-Redux-dev").to(dtype=dtype).to("cuda")



###   example  #####
ref_dir='./examples/ref_image'
ref_mask_dir='./examples/ref_mask'
image_dir='./examples/source_image'
image_mask_dir='./examples/source_mask'

ref_list=[os.path.join(ref_dir,file) for file in os.listdir(ref_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file ]
ref_list.sort()

ref_mask_list=[os.path.join(ref_mask_dir,file) for file in os.listdir(ref_mask_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file]
ref_mask_list.sort()

image_list=[os.path.join(image_dir,file) for file in os.listdir(image_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file ]
image_list.sort()

image_mask_list=[os.path.join(image_mask_dir,file) for file in os.listdir(image_mask_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file]
image_mask_list.sort()
###   example  #####



@spaces.GPU
def run_local(base_image, base_mask, reference_image, ref_mask, seed, base_mask_option, ref_mask_option, text_prompt):


    if base_mask_option == "Draw Mask":
        tar_image = base_image["background"]
        tar_mask = base_image["layers"][0]
    else:
        tar_image = base_image["background"]
        tar_mask = base_mask["background"]

    if ref_mask_option == "Draw Mask":
        ref_image = reference_image["background"]
        ref_mask = reference_image["layers"][0]
    elif ref_mask_option == "Upload with Mask":
        ref_image = reference_image["background"]
        ref_mask = ref_mask["background"]
    else:
        ref_image = reference_image["background"]
        ref_mask = get_mask(ref_image, text_prompt)

    tar_image = tar_image.convert("RGB")
    tar_mask = tar_mask.convert("L")
    ref_image = ref_image.convert("RGB")
    ref_mask = ref_mask.convert("L")

    return_ref_mask = ref_mask.copy()

    tar_image = np.asarray(tar_image)
    tar_mask = np.asarray(tar_mask)
    tar_mask = np.where(tar_mask > 128, 1, 0).astype(np.uint8)

    ref_image = np.asarray(ref_image)
    ref_mask = np.asarray(ref_mask)
    ref_mask = np.where(ref_mask > 128, 1, 0).astype(np.uint8)

    if tar_mask.sum() == 0:
        raise gr.Error('No mask for the background image.Please check mask button!')

    if ref_mask.sum() == 0:
        raise gr.Error('No mask for the reference image.Please check mask button!')

    ref_box_yyxx = get_bbox_from_mask(ref_mask)
    ref_mask_3 = np.stack([ref_mask,ref_mask,ref_mask],-1)
    masked_ref_image = ref_image * ref_mask_3 + np.ones_like(ref_image) * 255 * (1-ref_mask_3) 
    y1,y2,x1,x2 = ref_box_yyxx
    masked_ref_image = masked_ref_image[y1:y2,x1:x2,:]
    ref_mask = ref_mask[y1:y2,x1:x2] 
    ratio = 1.3
    masked_ref_image, ref_mask = expand_image_mask(masked_ref_image, ref_mask, ratio=ratio)


    masked_ref_image = pad_to_square(masked_ref_image, pad_value = 255, random = False) 

    kernel = np.ones((7, 7), np.uint8)
    iterations = 2
    tar_mask = cv2.dilate(tar_mask, kernel, iterations=iterations)

    # zome in
    tar_box_yyxx = get_bbox_from_mask(tar_mask)
    tar_box_yyxx = expand_bbox(tar_mask, tar_box_yyxx, ratio=1.2)

    tar_box_yyxx_crop =  expand_bbox(tar_image, tar_box_yyxx, ratio=2)    #1.2 1.6
    tar_box_yyxx_crop = box2squre(tar_image, tar_box_yyxx_crop) # crop box
    y1,y2,x1,x2 = tar_box_yyxx_crop


    old_tar_image = tar_image.copy()
    tar_image = tar_image[y1:y2,x1:x2,:]
    tar_mask = tar_mask[y1:y2,x1:x2]

    H1, W1 = tar_image.shape[0], tar_image.shape[1]
    # zome in


    tar_mask = pad_to_square(tar_mask, pad_value=0)
    tar_mask = cv2.resize(tar_mask, size)

    masked_ref_image = cv2.resize(masked_ref_image.astype(np.uint8), size).astype(np.uint8)
    pipe_prior_output = redux(Image.fromarray(masked_ref_image))


    tar_image = pad_to_square(tar_image, pad_value=255)

    H2, W2 = tar_image.shape[0], tar_image.shape[1]

    tar_image = cv2.resize(tar_image, size)
    diptych_ref_tar = np.concatenate([masked_ref_image, tar_image], axis=1)


    tar_mask = np.stack([tar_mask,tar_mask,tar_mask],-1)
    mask_black = np.ones_like(tar_image) * 0
    mask_diptych = np.concatenate([mask_black, tar_mask], axis=1)


    diptych_ref_tar = Image.fromarray(diptych_ref_tar)
    mask_diptych[mask_diptych == 1] = 255
    mask_diptych = Image.fromarray(mask_diptych)



    generator = torch.Generator("cuda").manual_seed(seed)
    edited_image = pipe(
        image=diptych_ref_tar,
        mask_image=mask_diptych,
        height=mask_diptych.size[1],
        width=mask_diptych.size[0],
        max_sequence_length=512,
        generator=generator,
        **pipe_prior_output, 
    ).images[0]



    width, height = edited_image.size
    left = width // 2
    right = width
    top = 0
    bottom = height
    edited_image = edited_image.crop((left, top, right, bottom))


    edited_image = np.array(edited_image)
    edited_image = crop_back(edited_image, old_tar_image, np.array([H1, W1, H2, W2]), np.array(tar_box_yyxx_crop)) 
    edited_image = Image.fromarray(edited_image)

    if ref_mask_option != "Label to Mask":
        return [edited_image]
    else:
        return [return_ref_mask, edited_image]   

def update_ui(option):
    if option == "Draw Mask":
        return gr.update(visible=False), gr.update(visible=True)
    else:
        return gr.update(visible=True), gr.update(visible=False)


with gr.Blocks() as demo:


    gr.Markdown("# Insert-Anything")
    gr.Markdown("### Make sure to select the correct mask button!!")


    with gr.Row():
        with gr.Column(scale=1):
            with gr.Row():
                base_image = gr.ImageEditor(label="Background Image", sources="upload", type="pil", brush=gr.Brush(colors=["#FFFFFF"],default_size = 30,color_mode = "fixed"),
                                    layers = False,
                                    interactive=True)

                base_mask = gr.ImageEditor(label="Background Mask", sources="upload", type="pil", layers = False, brush=False, eraser=False)

            with gr.Row():
                base_mask_option = gr.Radio(["Draw Mask", "Upload with Mask"], label="Background Mask Input Option", value="Upload with Mask")

            with gr.Row():
                ref_image = gr.ImageEditor(label="Reference Image", sources="upload", type="pil", brush=gr.Brush(colors=["#FFFFFF"],default_size = 30,color_mode = "fixed"),
                                    layers = False,
                                    interactive=True)
                                    
                ref_mask = gr.ImageEditor(label="Reference Mask", sources="upload", type="pil", layers = False, brush=False, eraser=False)

            with gr.Row():
                ref_mask_option = gr.Radio(["Draw Mask", "Upload with Mask", "Label to Mask"], label="Reference Mask Input Option", value="Upload with Mask")

            with gr.Row():
                text_prompt = gr.Textbox(label="Label")

        with gr.Column(scale=1):
            baseline_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery", height=800, columns=1)
            with gr.Accordion("Advanced Option", open=True):
                seed = gr.Slider(label="Seed", minimum=-1, maximum=999999999, step=1, value=666)
                gr.Markdown("### Guidelines")
                gr.Markdown(" Users can try using different seeds. For example, seeds like 42 and 123456 may produce different effects.")
                gr.Markdown(" Label to Mask means generating a mask by simply inputting a label.")

    run_local_button = gr.Button(value="Run")

    # #### example #####
    num_examples = len(image_list)
    for i in range(num_examples):
        with gr.Row():
            if i == 0:
                gr.Examples([image_list[i]], inputs=[base_image], label="Examples - Background Image", examples_per_page=1)
                gr.Examples([image_mask_list[i]], inputs=[base_mask], label="Examples - Background Mask", examples_per_page=1)
                gr.Examples([ref_list[i]], inputs=[ref_image], label="Examples - Reference Object", examples_per_page=1)
                gr.Examples([ref_mask_list[i]], inputs=[ref_mask], label="Examples - Reference Mask", examples_per_page=1)
            else:
                gr.Examples([image_list[i]], inputs=[base_image], examples_per_page=1, label="")
                gr.Examples([image_mask_list[i]], inputs=[base_mask], examples_per_page=1, label="")
                gr.Examples([ref_list[i]], inputs=[ref_image], examples_per_page=1, label="")
                gr.Examples([ref_mask_list[i]], inputs=[ref_mask], examples_per_page=1, label="")
        if i < num_examples - 1:
            gr.HTML("<hr>")
    # #### example #####

    run_local_button.click(fn=run_local,
                            inputs=[base_image, base_mask, ref_image, ref_mask, seed, base_mask_option, ref_mask_option, text_prompt],
                            outputs=[baseline_gallery]
                            )
demo.launch()