import tempfile import time from collections.abc import Sequence from typing import Any, cast import gradio as gr import numpy as np import pillow_heif import spaces import torch from gradio_image_annotation import image_annotator from gradio_imageslider import ImageSlider from PIL import Image from pymatting.foreground.estimate_foreground_ml import estimate_foreground_ml from refiners.fluxion.utils import no_grad from refiners.solutions import BoxSegmenter from transformers import GroundingDinoForObjectDetection, GroundingDinoProcessor BoundingBox = tuple[int, int, int, int] pillow_heif.register_heif_opener() pillow_heif.register_avif_opener() device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # weird dance because ZeroGPU segmenter = BoxSegmenter(device="cpu") segmenter.device = device segmenter.model = segmenter.model.to(device=segmenter.device) gd_model_path = "IDEA-Research/grounding-dino-base" gd_processor = GroundingDinoProcessor.from_pretrained(gd_model_path) gd_model = GroundingDinoForObjectDetection.from_pretrained(gd_model_path, torch_dtype=torch.float32) gd_model = gd_model.to(device=device) # type: ignore assert isinstance(gd_model, GroundingDinoForObjectDetection) def bbox_union(bboxes: Sequence[list[int]]) -> BoundingBox | None: if not bboxes: return None for bbox in bboxes: assert len(bbox) == 4 assert all(isinstance(x, int) for x in bbox) return ( min(bbox[0] for bbox in bboxes), min(bbox[1] for bbox in bboxes), max(bbox[2] for bbox in bboxes), max(bbox[3] for bbox in bboxes), ) def corners_to_pixels_format(bboxes: torch.Tensor, width: int, height: int) -> torch.Tensor: x1, y1, x2, y2 = bboxes.round().to(torch.int32).unbind(-1) return torch.stack((x1.clamp_(0, width), y1.clamp_(0, height), x2.clamp_(0, width), y2.clamp_(0, height)), dim=-1) def gd_detect(img: Image.Image, prompt: str) -> BoundingBox | None: assert isinstance(gd_processor, GroundingDinoProcessor) # Grounding Dino expects a dot after each category. inputs = gd_processor(images=img, text=f"{prompt}.", return_tensors="pt").to(device=device) with no_grad(): outputs = gd_model(**inputs) width, height = img.size results: dict[str, Any] = gd_processor.post_process_grounded_object_detection( outputs, inputs["input_ids"], target_sizes=[(height, width)], )[0] assert "boxes" in results and isinstance(results["boxes"], torch.Tensor) bboxes = corners_to_pixels_format(results["boxes"].cpu(), width, height) return bbox_union(bboxes.numpy().tolist()) def apply_mask( img: Image.Image, mask_img: Image.Image, defringe: bool = True, ) -> Image.Image: assert img.size == mask_img.size img = img.convert("RGB") mask_img = mask_img.convert("L") if defringe: # Mitigate edge halo effects via color decontamination rgb, alpha = np.asarray(img) / 255.0, np.asarray(mask_img) / 255.0 foreground = cast(np.ndarray[Any, np.dtype[np.uint8]], estimate_foreground_ml(rgb, alpha)) img = Image.fromarray((foreground * 255).astype("uint8")) result = Image.new("RGBA", img.size) result.paste(img, (0, 0), mask_img) return result @spaces.GPU def _gpu_process( img: Image.Image, prompt: str | BoundingBox | None, ) -> tuple[Image.Image, BoundingBox | None, list[str]]: # Because of ZeroGPU shenanigans, we need a *single* function with the # `spaces.GPU` decorator that *does not* contain postprocessing. time_log: list[str] = [] if isinstance(prompt, str): t0 = time.time() bbox = gd_detect(img, prompt) time_log.append(f"detect: {time.time() - t0}") if not bbox: print(time_log[0]) raise gr.Error("No object detected") else: bbox = prompt t0 = time.time() mask = segmenter(img, bbox) time_log.append(f"segment: {time.time() - t0}") return mask, bbox, time_log def _process( img: Image.Image, prompt: str | BoundingBox | None, ) -> tuple[tuple[Image.Image, Image.Image], gr.DownloadButton]: # enforce max dimensions for pymatting performance reasons if img.width > 2048 or img.height > 2048: orig_res = max(img.width, img.height) img.thumbnail((2048, 2048)) if isinstance(prompt, tuple): x0, y0, x1, y1 = (int(x * 2048 / orig_res) for x in prompt) prompt = (x0, y0, x1, y1) mask, bbox, time_log = _gpu_process(img, prompt) t0 = time.time() masked_alpha = apply_mask(img, mask, defringe=True) time_log.append(f"crop: {time.time() - t0}") print(", ".join(time_log)) masked_rgb = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha) thresholded = mask.point(lambda p: 255 if p > 10 else 0) bbox = thresholded.getbbox() to_dl = masked_alpha.crop(bbox) temp = tempfile.NamedTemporaryFile(delete=False, suffix=".png") to_dl.save(temp, format="PNG") temp.close() return (img, masked_rgb), gr.DownloadButton(value=temp.name, interactive=True) def process_bbox(prompts: dict[str, Any]) -> tuple[tuple[Image.Image, Image.Image], gr.DownloadButton]: assert isinstance(img := prompts["image"], Image.Image) assert isinstance(boxes := prompts["boxes"], list) if len(boxes) == 1: assert isinstance(box := boxes[0], dict) bbox = tuple(box[k] for k in ["xmin", "ymin", "xmax", "ymax"]) else: assert len(boxes) == 0 bbox = None return _process(img, bbox) def on_change_bbox(prompts: dict[str, Any] | None): return gr.update(interactive=prompts is not None) def process_prompt(img: Image.Image, prompt: str) -> tuple[tuple[Image.Image, Image.Image], gr.DownloadButton]: return _process(img, prompt) def on_change_prompt(img: Image.Image | None, prompt: str | None): return gr.update(interactive=bool(img and prompt)) TITLE = """
Create high-quality HD cutouts for any object in your image with just a text prompt — no manual work required!
The object will be available on a transparent background, ready to paste elsewhere.
This space uses the
Finegrain Box Segmenter model,
trained with a mix of natural data curated by Finegrain and
synthetic data provided by Nfinite.
It is powered by Refiners, our open source micro-framework for simple foundation model adaptation.
If you enjoyed it, please consider starring Refiners on GitHub!