"""llmgw's own ComfyUI nodes, put in every replica by the gate. LLMGWPrepareForMesh is what an image-to-3D model wants of its picture, as TripoSR's own pipeline (tsr/utils.py) does it: the background taken away — by the picture's alpha where it has one, by rembg (MIT; u2net, Apache 2.0) where it does not — the object cropped to its bounds, set square in the middle at a ratio of the side, on grey. """ import numpy as np import torch from PIL import Image _sessions = {} def _rembg(img, model): try: import rembg except ImportError as e: # a model that wants it names it in its nodes' requirements raise RuntimeError("the picture has no alpha and rembg is not installed for this model: " "add rembg and onnxruntime to its custom node requirements") from e if model not in _sessions: _sessions[model] = rembg.new_session(model) return rembg.remove(img, session=_sessions[model]) def _fit(rgba, ratio): """tsr.utils.resize_foreground: crop to the alpha's bounds, pad square, pad again so the object is ratio of the side.""" a = np.array(rgba) ys, xs = np.where(a[..., 3] > 0) if len(ys) == 0: raise RuntimeError("nothing is left of the picture once its background is taken away") fg = a[ys.min():ys.max() + 1, xs.min():xs.max() + 1] size = max(fg.shape[0], fg.shape[1]) ph0, pw0 = (size - fg.shape[0]) // 2, (size - fg.shape[1]) // 2 sq = np.pad(fg, ((ph0, size - fg.shape[0] - ph0), (pw0, size - fg.shape[1] - pw0), (0, 0))) new = int(size / ratio) ph0, pw0 = (new - size) // 2, (new - size) // 2 return np.pad(sq, ((ph0, new - size - ph0), (pw0, new - size - pw0), (0, 0))) class LLMGWPrepareForMesh: @classmethod def INPUT_TYPES(cls): return { "required": { "image": ("IMAGE",), "foreground_ratio": ("FLOAT", {"default": 0.85, "min": 0.3, "max": 1.0, "step": 0.01}), "background": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), "remove_background": (["auto", "always", "never"],), "rembg_model": (["u2net", "isnet-general-use", "u2netp"],), }, "optional": {"mask": ("MASK",)}, } RETURN_TYPES = ("IMAGE",) FUNCTION = "prepare" CATEGORY = "llmgw" def prepare(self, image, foreground_ratio, background, remove_background, rembg_model, mask=None): out = [] for i in range(image.shape[0]): rgb = (image[i].cpu().numpy() * 255).clip(0, 255).astype(np.uint8) h, w = rgb.shape[:2] alpha = None # LoadImage's mask is 1 - alpha, and a 64x64 of zeros for a # picture with none. if mask is not None and remove_background != "always": m = mask[min(i, mask.shape[0] - 1)].cpu().numpy() if m.shape == (h, w) and m.max() > 0: alpha = ((1.0 - m) * 255).clip(0, 255).astype(np.uint8) if alpha is None and remove_background != "never": rgba = np.array(_rembg(Image.fromarray(rgb), rembg_model).convert("RGBA")) else: if alpha is None: alpha = np.full((h, w), 255, np.uint8) rgba = np.dstack([rgb, alpha]) sq = _fit(rgba, foreground_ratio).astype(np.float32) / 255.0 a = sq[..., 3:4] img = sq[..., :3] * a + (1.0 - a) * background out.append(torch.from_numpy(img)) # One size for the batch: the first picture's. side = out[0].shape[0] out = [o if o.shape[0] == side else torch.from_numpy( np.array(Image.fromarray((o.numpy() * 255).astype(np.uint8)).resize((side, side)))).float() / 255.0 for o in out] return (torch.stack(out),) NODE_CLASS_MAPPINGS = {"LLMGWPrepareForMesh": LLMGWPrepareForMesh} NODE_DISPLAY_NAME_MAPPINGS = {"LLMGWPrepareForMesh": "Prepare a picture for a 3D model (llmgw)"}