dn04.py

# FaceDetailerStandalone_MIN_FIXED_FAST_EMBEDDED_SAM.py
# One-node Face Detailer (image-only) with fixed settings + embedded Ultralytics bbox detector + embedded SAM loader.
# - Output parity with Impact Pack Face Detailer at the same settings
# - No separate bbox-detector node; detector is cached/constructed internally
# - No separate SAM loader node; SAM is cached/constructed internally
# - Lightweight runtime overhead (cached imports, inference_mode, fused layers, TF32, FP16 on CUDA)

import os
from dataclasses import dataclass
from typing import List, Tuple, Optional

import warnings
warnings.filterwarnings("ignore")

# Silence OpenCV before importing it (env var) and after (setLogLevel)
os.environ["OPENCV_LOG_LEVEL"] = "ERROR"

import numpy as np
import torch
import comfy
from PIL import Image
import cv2

try:
    if hasattr(cv2, "setLogLevel"):
        try:
            lvl = cv2.LOG_LEVEL_ERROR if hasattr(cv2, "LOG_LEVEL_ERROR") else 3  # 3 == error
            cv2.setLogLevel(lvl)
        except Exception:
            pass
except Exception:
    pass

# ---------------- Fixed FaceDetailer settings (do not expose in UI) ----------------
# GUIDE_SIZE = 512
# GUIDE_SIZE_FOR_BBOX = True
# MAX_SIZE = 1024
# STEPS = 30
# CFG = 7.0
# SCHEDULER = "simple"
# DENOISE = 0.5
# FEATHER = 5
# NOISE_MASK = True
# FORCE_INPAINT = True
# BBOX_THRESHOLD = 0.5
# BBOX_DILATION = 10
# BBOX_CROP_FACTOR = 3.0
# DROP_SIZE = 10
# SAM_DETECTION_HINT = "center-1"
# SAM_DILATION = 0
# SAM_THRESHOLD = 0.93
# SAM_BBOX_EXPANSION = 0
# SAM_MASK_HINT_THRESHOLD = 0.7
# SAM_MASK_HINT_USE_NEGATIVE = "False"
# WILDCARD = ""
# CYCLE = 1
# INPAINT_MODEL = False
# NOISE_MASK_FEATHER = 20
# TILED_ENCODE = False
# TILED_DECODE = False
# ---------------------------------------------------------------------

# ---------------- Ultralytics / YOLO detector integration (embedded) ----------------

# Torch runtime perf switches
torch.backends.cudnn.benchmark = True  # autotune best conv algorithms
if torch.cuda.is_available():
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True
    try:
        torch.set_float32_matmul_precision("high")  # PyTorch 2.x
    except Exception:
        pass

# Optional Impact Pack interop (SEG type)
try:
    # If Impact Pack is installed, use its SEG to be perfectly compatible.
    from impact.core import SEG as _IMPACT_SEG  # type: ignore
    _USE_IMPACT_SEG = True
except Exception:
    _USE_IMPACT_SEG = False

    @dataclass
    class _LocalSEG:
        cropped_image: Optional[torch.Tensor]
        cropped_mask: np.ndarray  # 2D float32 [0..1]
        confidence: float
        crop_region: Tuple[int, int, int, int]  # (x1,y1,x2,y2)
        bbox: Tuple[int, int, int, int]         # (x1,y1,x2,y2)
        label: str
        control_net_wrapper: Optional[object] = None

SEG = _IMPACT_SEG if _USE_IMPACT_SEG else _LocalSEG

# ---------------------------------------------------------------------
# LOCAL ASSET PATHS (no hardcoded absolute paths)
# ---------------------------------------------------------------------

# Base directory of this node file (cross-platform, works on RunPod/ComfyUI)
BASE_DIR = os.path.dirname(os.path.abspath(__file__))

# Local YOLO model path inside this custom node folder
YOLO_MODEL_PATH = os.path.join(BASE_DIR, "assets", "face_yolov8m_salia.pt")
YOLO_IMGSZ = 640

# Local SAM checkpoint path inside this custom node folder
SAM_CKPT_PATH = os.path.join(BASE_DIR, "assets", "sam_vit_b_01ec64_salia.pth")

# Cached instances (process-local)
_CACHED_YOLO_MODEL = None
_CACHED_ULTRA_DETECTOR = None


def _tensor_to_pil(image: torch.Tensor) -> Image.Image:
    # image: [1, H, W, 3], float(0..1)
    img = image[0].detach().cpu().clamp(0, 1).numpy()
    img = (img * 255.0).round().astype(np.uint8)  # (H, W, 3) RGB
    return Image.fromarray(img, mode="RGB")


def _make_crop_region(w: int, h: int, bbox_xyxy, crop_factor: float) -> Tuple[int, int, int, int]:
    x1, y1, x2, y2 = map(int, bbox_xyxy)
    cx = (x1 + x2) * 0.5
    cy = (y1 + y2) * 0.5
    bw = (x2 - x1)
    bh = (y2 - y1)
    new_w = max(1, int(bw * crop_factor))
    new_h = max(1, int(bh * crop_factor))
    # center to image
    nx1 = int(max(0, round(cx - new_w * 0.5)))
    ny1 = int(max(0, round(cy - new_h * 0.5)))
    nx2 = int(min(w, nx1 + new_w))
    ny2 = int(min(h, ny1 + new_h))
    # clamp again
    nx1 = max(0, min(nx1, w - 1))
    ny1 = max(0, min(ny1, h - 1))
    nx2 = max(nx1 + 1, min(nx2, w))
    ny2 = max(ny1 + 1, min(ny2, h))
    return (nx1, ny1, nx2, ny2)


def _crop_tensor_image(image: torch.Tensor, crop: Tuple[int, int, int, int]) -> torch.Tensor:
    # image: [1,H,W,3]; crop: (x1,y1,x2,y2)
    x1, y1, xb, yb = crop
    return image[:, y1:yb, x1:xb, :].contiguous()


def _crop_ndarray(mask: np.ndarray, crop: Tuple[int, int, int, int]) -> np.ndarray:
    # mask: [H,W] float/bool/uint8; crop: (x1,y1,x2,y2)
    x1, y1, xb, yb = crop
    return mask[int(y1):int(yb), int(x1):int(xb)]


def _dilate_masks(segmasks: List[Tuple[np.ndarray, np.ndarray, float]], factor: int):
    if factor == 0 or not segmasks:
        return segmasks
    k = abs(int(factor))
    if k < 1:
        return segmasks
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (k, k))
    do_dilate = factor > 0
    out = []
    for (bbox, m, conf) in segmasks:
        u8 = (m * 255.0).astype(np.uint8) if m.dtype != np.uint8 else m
        d = cv2.dilate(u8, kernel, iterations=1) if do_dilate else cv2.erode(u8, kernel, iterations=1)
        out.append((bbox, d.astype(np.float32) / 255.0, conf))
    return out


def _combine_masks(segmasks: List[Tuple[np.ndarray, np.ndarray, float]]) -> Optional[torch.Tensor]:
    if not segmasks:
        return None
    h = segmasks[0][1].shape[0]
    w = segmasks[0][1].shape[1]
    acc = np.zeros((h, w), dtype=np.uint8)
    for _, m, _ in segmasks:
        u8 = (m * 255.0).astype(np.uint8) if m.dtype != np.uint8 else m
        acc = cv2.bitwise_or(acc, u8)
    return torch.from_numpy(acc.astype(np.float32) / 255.0)  # [H,W], float32 0..1 CPU


def _pick_device_str(user_device: str = "") -> str:
    if user_device:
        return user_device
    return "cuda" if torch.cuda.is_available() else "cpu"


@torch.inference_mode()
def _inference_bbox(model, image_pil: Image.Image, confidence: float = 0.3, device: str = ""):
    """
    Returns results = [labels(str), bboxes(xyxy), segms(full-image bool masks), conf(float)]
    For bbox models, segm "masks" are rectangles from the boxes (Subpack parity).
    """
    pred = model(
        image_pil,
        conf=confidence,
        device=_pick_device_str(device),
        verbose=False,
        imgsz=YOLO_IMGSZ,  # fixed size can be faster
    )

    p0 = pred[0]
    boxes = p0.boxes
    bboxes = boxes.xyxy.detach().cpu().numpy()  # (N,4) float, xyxy

    W, H = image_pil.size
    segms = []
    for x0, y0, x1, y1 in bboxes:
        m = np.zeros((H, W), np.uint8)
        cv2.rectangle(m, (int(x0), int(y0)), (int(x1), int(y1)), 255, -1)
        segms.append(m.astype(bool))

    if bboxes.shape[0] == 0:
        return [[], [], [], []]

    results = [[], [], [], []]
    names = p0.names
    for i, (bbox, segm) in enumerate(zip(bboxes, segms)):
        cls_i = int(boxes.cls[i].item())
        results[0].append(names[cls_i])
        results[1].append(bbox)
        results[2].append(segm)
        results[3].append(float(boxes.conf[i].item()))
    return results


def _create_segmasks(results):
    bboxes = results[1]
    segms = results[2]
    confs = results[3]
    out = []
    for i in range(len(segms)):
        out.append((bboxes[i], segms[i].astype(np.float32), confs[i]))
    return out


class UltraBBoxDetector:
    def __init__(self, yolo_model):
        self.bbox_model = yolo_model

    def detect(self, image, threshold, dilation, crop_factor, drop_size=1, detailer_hook=None):
        drop_size = max(int(drop_size), 1)
        detected = _inference_bbox(self.bbox_model, _tensor_to_pil(image), threshold)
        segmasks = _create_segmasks(detected)
        if int(dilation) != 0:
            segmasks = _dilate_masks(segmasks, int(dilation))

        H = int(image.shape[1])
        W = int(image.shape[2])
        items = []
        for (bbox_xyxy, full_mask, conf), label in zip(segmasks, detected[0]):
            x1, y1, x2, y2 = map(int, bbox_xyxy)
            if (x2 - x1) > drop_size and (y2 - y1) > drop_size:
                crop_region = _make_crop_region(W, H, (x1, y1, x2, y2), float(crop_factor))
                if detailer_hook is not None and hasattr(detailer_hook, "post_crop_region"):
                    crop_region = detailer_hook.post_crop_region(W, H, (x1, y1, x2, y2), crop_region)

                cropped_image = _crop_tensor_image(image, crop_region)
                cropped_mask = _crop_ndarray(full_mask, crop_region).astype(np.float32)
                items.append(SEG(cropped_image, cropped_mask, float(conf), crop_region, (x1, y1, x2, y2), str(label), None))

        segs = ((H, W), items)
        if detailer_hook is not None and hasattr(detailer_hook, "post_detection"):
            segs = detailer_hook.post_detection(segs)
        return segs

    def detect_combined(self, image, threshold, dilation):
        detected = _inference_bbox(self.bbox_model, _tensor_to_pil(image), threshold)
        segmasks = _create_segmasks(detected)
        if int(dilation) != 0:
            segmasks = _dilate_masks(segmasks, int(dilation))
        return _combine_masks(segmasks)

    def setAux(self, x):
        # kept for signature parity
        pass


def _load_ultralytics_model(model_path: str):
    # Import here so that module import doesn't hard-fail if ultralytics is missing
    try:
        from ultralytics import YOLO
    except Exception as e:
        raise RuntimeError(
            "[FaceDetailerStandalone] The 'ultralytics' package is required for the embedded bbox detector.\n"
            "Install in your ComfyUI python:  python -m pip install --upgrade ultralytics"
        ) from e

    if not os.path.isfile(model_path):
        raise FileNotFoundError(
            "[FaceDetailerStandalone] Embedded YOLO model file not found.\n"
            f"Expected at: {model_path}\n"
            "Please place 'face_yolov8m_salia.pt' in the 'assets' folder next to this node."
        )

    yolo = YOLO(model_path)

    # One-time graph/model optimizations
    try:
        dev = _pick_device_str()
        try:
            yolo.to(dev)  # newer Ultralytics
        except Exception:
            yolo.model.to(dev)  # older versions
    except Exception:
        pass

    # Fuse Conv+BN where possible (small speedup)
    try:
        yolo.fuse()
    except Exception:
        pass

    # Use half precision weights on CUDA (big win; safe for inference)
    try:
        if torch.cuda.is_available():
            yolo.model.half()
    except Exception:
        pass

    return yolo


def _get_embedded_detector():
    global _CACHED_YOLO_MODEL, _CACHED_ULTRA_DETECTOR
    if _CACHED_ULTRA_DETECTOR is not None:
        return _CACHED_ULTRA_DETECTOR
    if _CACHED_YOLO_MODEL is None:
        _CACHED_YOLO_MODEL = _load_ultralytics_model(YOLO_MODEL_PATH)
    _CACHED_ULTRA_DETECTOR = UltraBBoxDetector(_CACHED_YOLO_MODEL)
    return _CACHED_ULTRA_DETECTOR

# ---------------- Embedded SAM loader (GPU-only, hardcoded path, reuse one predictor) ----------------
# Matches your SAMLoaderStandalone design, but embedded + cached.


def _to_numpy_rgb(image_tensor):
    """
    Comfy 'IMAGE' is NHWC in [0..1]. Convert to uint8 HxWx3 RGB numpy.
    Accepts torch.Tensor (NHWC) or numpy already in HWC.
    """
    if isinstance(image_tensor, torch.Tensor):
        img = image_tensor
        if img.dim() == 4 and img.shape[0] == 1:
            img = img[0]
        img = (img.clamp(0, 1) * 255.0).to(torch.uint8).cpu().numpy()  # HWC
        return img
    elif isinstance(image_tensor, np.ndarray):
        if image_tensor.dtype != np.uint8:
            img = np.clip(image_tensor, 0, 255).astype(np.uint8)
        else:
            img = image_tensor
        return img
    else:
        raise TypeError(f"Unsupported image type for SAM: {type(image_tensor)}")


class _SAMWrapperGPUOnlyFast:
    """
    FaceDetailer-compatible wrapper:
      - Stays on CUDA
      - Reuses a single SamPredictor
      - predict(image, points, plabs, bbox, threshold) -> list[HxW float32 CPU masks]
    """
    def __init__(self, model):
        self.model = model
        dev = comfy.model_management.get_torch_device()
        if "cuda" not in str(dev).lower():
            raise RuntimeError(
                f"[FaceDetailerStandalone] GPU-only SAM: CUDA device not available (got '{dev}')."
            )
        self._device = dev
        self.model.to(self._device).eval()
        # Lazy import for segment_anything predictor
        from segment_anything import SamPredictor  # type: ignore
        # Reuse one predictor instance (cheaper than re-creating every call)
        self._predictor = SamPredictor(self.model)

    def prepare_device(self):
        if "cuda" not in str(self._device).lower():
            raise RuntimeError("[FaceDetailerStandalone] CUDA device lost/unavailable for SAM.")

    def release_device(self):
        # GPU-only; keep on GPU (no-op)
        pass

    @torch.inference_mode()
    def predict(self, image, points, plabs, bbox, threshold: float):
        """
        image: Comfy IMAGE (NHWC, [0..1]) or numpy
        points: list[[x,y], ...] or None
        plabs:  list[int] (1=fg, 0=bg) or None
        bbox:   [x1,y1,x2,y2] or None
        threshold: float in [0..1]
        returns: list of HxW float32 CPU masks (0/1)
        """
        self.prepare_device()

        np_img = _to_numpy_rgb(image)
        # Some builds call set_image(img, "RGB"); accept both signatures.
        try:
            self._predictor.set_image(np_img, "RGB")
        except TypeError:
            self._predictor.set_image(np_img)

        pc = np.array(points, dtype=np.float32) if points else None
        pl = np.array(plabs, dtype=np.int32) if plabs else None
        bx = np.array(bbox, dtype=np.float32) if bbox is not None else None

        # Keep provided behavior: multimask_output=False
        masks, scores, _ = self._predictor.predict(
            point_coords=pc,
            point_labels=pl,
            box=bx,
            multimask_output=False
        )

        out = []
        if masks is not None and scores is not None:
            for m, s in zip(masks, scores):
                if float(s) >= float(threshold):
                    if isinstance(m, torch.Tensor):
                        t = m.to(torch.float32).cpu()
                    else:
                        t = torch.from_numpy(m.astype(np.float32)).cpu()
                    out.append(t)
        return out


# Cache for SAM
_CACHED_SAM_MODEL = None


def _get_embedded_sam():
    """Load SAM vit_b from SAM_CKPT_PATH and attach GPU-only fast wrapper, cached."""
    global _CACHED_SAM_MODEL
    if _CACHED_SAM_MODEL is not None:
        return _CACHED_SAM_MODEL

    if not os.path.isfile(SAM_CKPT_PATH):
        raise FileNotFoundError(
            f"[FaceDetailerStandalone] SAM checkpoint not found:\n  {SAM_CKPT_PATH}\n"
            f"Place 'sam_vit_b_01ec64_salia.pth' in the 'assets' folder next to this node."
        )

    # Import here to avoid module import failure at file load time
    try:
        from segment_anything import sam_model_registry  # type: ignore
    except Exception as e:
        raise RuntimeError(
            "[FaceDetailerStandalone] 'segment_anything' is not installed for embedded SAM. "
            "Install in your Comfy python, e.g.:  python -m pip install "
            "git+https://github.com/facebookresearch/segment-anything"
        ) from e

    # Fixed to vit_b (matches 'sam_vit_b_01ec64' weights)
    sam = sam_model_registry['vit_b'](checkpoint=SAM_CKPT_PATH)
    sam.eval()  # ensure eval mode

    # Attach GPU-only, faster wrapper
    wrapper = _SAMWrapperGPUOnlyFast(sam)
    sam.sam_wrapper = wrapper

    _CACHED_SAM_MODEL = sam
    return _CACHED_SAM_MODEL

# ---------------- Impact Pack Face Detailer binding ----------------
_ENHANCE_FACE = None
_IMPORT_ERR = None
try:
    from impact.impact_pack import FaceDetailer as _FD
    _ENHANCE_FACE = _FD.enhance_face
except Exception as _e:
    _IMPORT_ERR = _e
    _ENHANCE_FACE = None

# ---------------- Single public node ----------------
class dn_04:
    @classmethod
    def INPUT_TYPES(cls):
        # Only essential, connectable parts remain editable. (No bbox or SAM inputs.)
        return {
            "required": {
                "image": ("IMAGE",),
                "model": ("MODEL", {"tooltip": "If `ImpactDummyInput` is connected to model, inference is skipped."}),
                "clip": ("CLIP",),
                "vae": ("VAE",),

                # Keep sampler selectable; all other knobs are fixed
                "sampler_name": (comfy.samplers.KSampler.SAMPLERS,),

                # Conditioning stays connectable
                "positive": ("CONDITIONING",),
                "negative": ("CONDITIONING",),

                # Keep seed editable but fixed after generate for reproducibility
                "seed": ("INT", {
                    "default": 0,
                    "min": 0,
                    "max": 0xffffffffffffffff,
                    "step": 1,
                    "control_after_generate": "fixed",
                }),
            },
            "optional": {
                # No external SAM input; embedded
            }
        }

    RETURN_TYPES = ("IMAGE",)
    RETURN_NAMES = ("image",)
    FUNCTION = "doit"
    CATEGORY = "ImpactPack/Standalone"
    DESCRIPTION = (
        "Face Detailer with requested parameters hardcoded (non-editable), "
        "and embedded Ultralytics face bbox detector + embedded SAM (no external input nodes). "
        "Optimized call path (cached imports + inference_mode) for lower overhead; "
        "results identical to Impact Pack Face Detailer at the same settings."
    )

    def doit(
        self,
        image, model, clip, vae,
        sampler_name,
        positive, negative,
        seed,
    ):
        if _ENHANCE_FACE is None:
            raise RuntimeError(
                "ComfyUI-Impact-Pack is required for Face Detailer logic. "
                "Please install/enable ComfyUI-Impact-Pack."
            ) from _IMPORT_ERR

        # Embedded detector & SAM (cached)
        bbox_detector = _get_embedded_detector()
        sam_model_opt = _get_embedded_sam()

        enhance = _ENHANCE_FACE

        # Determine batch size safely
        B = image.shape[0] if (hasattr(image, "shape") and image.ndim == 4) else 1

        # No autograd, faster kernel choices, identical math for inference
        with torch.inference_mode():
            if B == 1:
                # Fast-path for single image (avoid list + cat)
                single = image[0] if image.ndim == 4 else image  # [H,W,C]
                enhanced_img, _, _, _, _ = enhance(
                    single.unsqueeze(0),  # -> [1,H,W,C]
                    model, clip, vae,
                    512, True, 1024,          # guide_size, guide_for_bbox, max_size
                    seed, 38, 7.0,            # steps, cfg
                    sampler_name, "simple",   # scheduler name
                    positive, negative,
                    0.4, 5, True, True,       # denoise, feather, noise_mask, force_inpaint
                    0.5, 10, 3.0,             # bbox_threshold, bbox_dilation, bbox_crop_factor
                    "center-1", 0, 0.93, 0,   # sam_detection_hint, sam_dilation, sam_threshold, sam_bbox_expansion
                    0.7, "False",             # sam_mask_hint_threshold, sam_mask_hint_use_negative
                    10, bbox_detector,        # drop_size, bbox_detector
                    # Internals not exposed (kept fixed/None)
                    segm_detector=None, sam_model_opt=sam_model_opt,
                    wildcard_opt="", detailer_hook=None,
                    refiner_ratio=None, refiner_model=None, refiner_clip=None,
                    refiner_positive=None, refiner_negative=None,
                    cycle=1, inpaint_model=False,
                    noise_mask_feather=20,
                    scheduler_func_opt=None,
                    tiled_encode=False, tiled_decode=False,
                )
                return (enhanced_img,)

            # Batch of images; per-frame process with seed+i
            out_imgs = []
            for i, single in enumerate(image.unbind(0)):
                enhanced_img, _, _, _, _ = enhance(
                    single.unsqueeze(0),  # [1,H,W,C]
                    model, clip, vae,
                    512, True, 1024,
                    seed + i, 30, 7.0,
                    sampler_name, "simple",
                    positive, negative,
                    0.5, 5, True, True,
                    0.5, 10, 3.0,
                    "center-1", 0, 0.93, 0,
                    0.7, "False",
                    10, bbox_detector,
                    segm_detector=None, sam_model_opt=sam_model_opt,
                    wildcard_opt="", detailer_hook=None,
                    refiner_ratio=None, refiner_model=None, refiner_clip=None,
                    refiner_positive=None, refiner_negative=None,
                    cycle=1, inpaint_model=False,
                    noise_mask_feather=20,
                    scheduler_func_opt=None,
                    tiled_encode=False, tiled_decode=False,
                )
                out_imgs.append(enhanced_img)

            return (torch.cat(out_imgs, dim=0),)


NODE_CLASS_MAPPINGS = {
    "dn_04": dn_04,
}
NODE_DISPLAY_NAME_MAPPINGS = {
    "dn_04": "dn_04",
}