roboflow/supervision · error · ValueError

addWeighted fallback only supports the default output depth;

Error message

addWeighted fallback only supports the default output depth; unsupported dtype: {dtype}

What it means

OpenCV's `addWeighted` accepts a `dtype` (ddepth) argument to control the output type. The fallback at src/supervision/_cv2/_image.py:80 only supports the default behavior (output dtype = src1's dtype, i.e. CV_16F/CV_32F promotion aside, it casts back like the source), so any explicit dtype other than the sentinel -1 raises.

Source

Thrown at src/supervision/_cv2/_image.py:80

        fill_value = fill.reshape((1, 1, -1))

    result = np.full(shape, fill_value, dtype=image.dtype)
    result[top : top + height, left : left + width] = image
    return result


def _add_weighted(
    src1: npt.NDArray[Any],
    alpha: float,
    src2: npt.NDArray[Any],
    beta: float,
    gamma: float,
    dst: npt.NDArray[Any] | None = None,
    dtype: int | None = None,
) -> npt.NDArray[Any]:
    """Blend two arrays with OpenCV-compatible saturation and optional mutation."""
    if dtype is not None and dtype != -1:
        raise ValueError(
            "addWeighted fallback only supports the default output depth; "
            f"unsupported dtype: {dtype}"
        )
    if src1.shape != src2.shape:
        raise ValueError("addWeighted inputs must have equal shapes")
    result = _cast_array_like_opencv(
        src1.astype(np.float64) * alpha + src2.astype(np.float64) * beta + gamma,
        src1.dtype,
    )
    if dst is not None:
        dst[...] = result
        return dst
    return result


def _convert_scale_abs(
    image: npt.NDArray[Any], alpha: float = 1, beta: float = 0
) -> npt.NDArray[np.uint8]:

View on GitHub (pinned to 7f254d9784)

Solutions

1. Drop the `dtype` argument and let the output inherit src1's dtype
2. Do the blend manually at the desired precision: `(a.astype(np.float64)*alpha + b.astype(np.float64)*beta + gamma).astype(np.float32)`
3. Install `opencv-python` if ddepth control is required

Example fix

// before
out = cv2.addWeighted(a, 0.7, b, 0.3, 0.0, dtype=cv2.CV_32F)

// after
out = (a.astype(np.float64) * 0.7 + b.astype(np.float64) * 0.3).astype(np.float32)

import numpy as np

def blend_portable(a, alpha: float, b, beta: float, gamma: float = 0.0, dtype=None):
    """addWeighted without the ddepth argument; optional manual output dtype."""
    out = a.astype(np.float64) * alpha + b.astype(np.float64) * beta + gamma
    return out.astype(dtype) if dtype is not None else out.astype(a.dtype)

When it happens

Trigger: Calling `cv2.addWeighted(a, alpha, b, beta, gamma, dtype=cv2.CV_16U)` (or CV_32F etc.) on the fallback backend — typical in HDR-style blending or when a higher-precision accumulator is wanted to avoid saturation.

Common situations: Image-blending code written against real OpenCV requesting a wider output depth, run in an environment without opencv-python.

Related errors

• Unsupported color conversion code: {code}
• Only BORDER_CONSTANT is supported by the fallback
• Detection annotation for image {image_path} contains non-int
• Detections class_id must be an integer for Pascal VOC export
• Detections class_id must be an integer for YOLO export, got

AI-assisted analysis of roboflow/supervision@7f254d9784 (2026-08-15). Data as JSON: /api/errors/8b653b612f512b6f. Report an issue: GitHub.