DiagSplat · local multiradius boundary critic

Frame 0030 · 四个失败位置 × 四条候选线 → 新 mask

每个 block 的 RGB 和 residual 均作为独立 512×512 PNG 输入。 C1 白=+64，C2 洋红=+128，C3 黄=+192，C4 薄荷绿=+256 px。 选择最小的完整覆盖候选后，以 Gaussian 空间权重平滑组合局部半径，避免方形 crop 接缝。

Residual 仍是有偏辅助：轮廓外 residual 可以否决候选， 但低 residual 不能证明覆盖。最终组合使用 RGB-first 的选择。

Block 01 · C1 · +64 px

Confidence 0.87. C1 encloses the large blurred leaf, its cyan fringe, and the adjacent flower branch with a visible buffer.

Separate critic input A · RGB · four candidate lines
Separate critic input B · biased residual · four candidate lines

Block 02 · C1 · +64 px

Confidence 0.88. The thin upward branch, clipped tip, blossoms, and defocus halo remain between the C1 boundaries; larger contours only absorb background.

Separate critic input A · RGB · four candidate lines
Separate critic input B · biased residual · four candidate lines

Block 03 · C2 · +128 px

Confidence 0.90. The broad cyan/green defocus halo from the foreground foliage crosses C1 near the upper edge; C2 is the first candidate that contains it.

Separate critic input A · RGB · four candidate lines
Separate critic input B · biased residual · four candidate lines

Block 04 · C2 · +128 px

Confidence 0.82. The diffuse foreground halo approaches and locally meets C1 along the lower-left edge, so the conservative next radius C2 is required.

Separate critic input A · RGB · four candidate lines
Separate critic input B · biased residual · four candidate lines

四个判断组合得到的新 mask

Hard mask fraction 19.648% → 26.514%；新增 6.865%。 下载最终 mask PNG

old-mask
new-mask
added-hard-region
spatially-smoothed-radius-field
new-mask-on-biased-residual
完整 critic prompt
Judge ONE frame-0030 boundary location at a time.

You receive two separate 512x512 PNG images for the same location:
A. true observed RGB;
B. biased Round-1 residual.
They are separate visual inputs, never a merged composite.

Four thin candidate hard-mask boundaries appear in both images:
C1 white/black = +64 px
C2 magenta/black = +128 px
C3 yellow/black = +192 px
C4 mint/black = +256 px

Choose the SMALLEST candidate that encloses every visible part of the local
transient object indicated by the current mask. Include thin tips, small
protrusions, translucent or defocused boundaries, motion/defocus halos, and
color/blur trails. The existing mask may under-cover the object. Do not expand
merely to include unrelated static background. If even C4 fails, return
INSUFFICIENT.

Residual construction: mean absolute RGB difference between dirty observed and
the Round-1 3DGS render, Gaussian sigma 1.8, clipped at 0.30, Turbo colormap.
The render was trained with an earlier undercovered mask and can reproduce
leaked transient pixels. Therefore coherent residual outside a line proves
that line insufficient, but low residual never proves coverage.

Return: selected candidate, radius, confidence, and one concise RGB-first
reason.