The paper presents an extension of learnable tree filter method (LTF). On the positive side, the proposed method looks reasonable and it leads to improved performance over the baselines. However, on the weak side, the proposed contributions are fairly incremental, and performance improvement is somewhat marginal (over LTF-v1 and other SOTA methods), and comparisons against some state-of-the-art methods are missing. Overall, given the strong support from three reviewers out of four, I would give a chance for the authors to address the issues. If accepted, the authors should revise the paper according to their rebuttal.