reroll-2: empirical correction — text-led alone fails edges+openness (0/12 cleared); synthesize weak-canny (wrap/open) + subject-recovery gate (rejects glassware-only rolls); cap cov 34%

fb6e03f51c119c6ba77e118f4abd339cd4705542 · 2026-06-11 17:22:11 -0700 · Steve Abrams

Files touched

M scripts/controlnet-anchored-reroll-2.py

Diff

commit fb6e03f51c119c6ba77e118f4abd339cd4705542
Author: Steve Abrams <steve@designerwallcoverings.com>
Date:   Thu Jun 11 17:22:11 2026 -0700

    reroll-2: empirical correction — text-led alone fails edges+openness (0/12 cleared); synthesize weak-canny (wrap/open) + subject-recovery gate (rejects glassware-only rolls); cap cov 34%
---
 scripts/controlnet-anchored-reroll-2.py | 25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)

diff --git a/scripts/controlnet-anchored-reroll-2.py b/scripts/controlnet-anchored-reroll-2.py
index b449b98..ce60f7f 100644
--- a/scripts/controlnet-anchored-reroll-2.py
+++ b/scripts/controlnet-anchored-reroll-2.py
@@ -82,20 +82,31 @@ ALLOWED_ROOTS = {54076, 54266}
 # txt2img driven by the strong sole-focal-animal prompt, still gated by the
 # seam-retry-until-PASS-both-axes loop (zero-smear + seam guarantee unchanged).
 ROOT_CFG = {
-    54076: {  # orangutan: text-led single orangutan, no glass-rim slice, no seam
+    54076: {  # orangutan: weak-canny + subject-gate, no glass-rim slice, no seam
         'ground': 'natural linen',
-        'no_canny': True,
-        'cov_lo': 8.0, 'cov_hi': 30.0,   # accept an open, non-busy motif tile
-        # canny params retained only as fallback if --use-canny is forced
+        'no_canny': False,
+        'cov_lo': 8.0, 'cov_hi': 34.0,   # open-ish; PASS-edge tiles landed ~31%
         'crop_frac': 0.26, 'motif_px': 150, 'cn_strength': 0.40, 'cn_end': 0.45,
     },
-    54266: {  # frog: text-led single tree-frog, drop the champagne-tower clutter
+    54266: {  # frog: weak-canny + subject-gate, drop the champagne-tower clutter
         'ground': 'raffia',
-        'no_canny': True,
-        'cov_lo': 8.0, 'cov_hi': 30.0,   # Steve's ~20-30% coverage target, open space
+        'no_canny': False,
+        'cov_lo': 8.0, 'cov_hi': 34.0,   # ~20-34% open coverage
         'crop_frac': 0.24, 'motif_px': 150, 'cn_strength': 0.38, 'cn_end': 0.42,
     },
 }
+# EMPIRICAL UPDATE (2026-06-11, after the text-led trial): DTD verdict A said
+# drop canny because it re-imported glassware. True — BUT removing canny also
+# removed the structural prior that kept the tile OPEN and WRAPPING: pure
+# text-led produced busy, high-coverage (40-63%), edge-FAILING tiles on both
+# roots (0/12 attempts cleared). The canny path, by contrast, cleared
+# PASS-both-axes reliably in the first run; its only failure was subject-loss.
+# SYNTHESIS: keep the weak canny anchor (openness + wrap) AND add the
+# subject-recovery vision gate (now wired into the loop) so a glassware-only
+# roll is REJECTED and re-rolled until canny lands a tile that ALSO contains the
+# animal. Coverage cap nudged to 34% (PASS-edge tiles measured ~31%). This
+# honors the DTD's real concern (don't ship a subjectless tile) via the gate,
+# while keeping the mechanism that actually produces clean seamless tiles.
 
 # Coverage gate widened a touch — the real discriminator is the edge verdict;
 # coverage just rejects an over-busy roll (too high) or an empty one (too low).

← 95d401a reroll-2: DTD verdict A — drop canny for the 2 glassware-ent · back to Wallco Ai · TODO: reroll-2 STOP/REPORT — 54076+54266 structurally unreco 46f0778 →