This problem does not happen in common test condition. It only happens when g_uiMaxCUDepth is equal to 0 and 1-pass ALF is applied.

When STAR_CROSS_SHAPES_LUMA is defined, the selected filter shape for training new coefficients may be wrong.

solution

One-line fix. In Line 4851, TencAdaptiveLoopFilter.cpp, add the following code.

#if STAR_CROSS_SHAPES_LUMA

copyALFParam(m_pcTempAlfParam, &cAlfParam);

#endif