[object Object]

← back to Wallco Ai

mid_seam_heal.py — $0 production fix for SDXL latent-grid leak

02f91e6f44d8923383a0a1210ef1829d16e77e04 · 2026-05-26 08:23:37 -0700 · Steve Abrams

Replaces the planned $293 Gemini-edit run on the 7,333 mids-only FAIL
cohort. PIL Gaussian-band-blur (BAND=12px, SIGMA=4.0) at H/2 and V/2
midlines collapses the 4x4 latent-block stitch discontinuity while
preserving composition everywhere else.

Validated on 20-sample experiment (committed e30db7df):
  20/20 improved · 9 → PASS · 11 → WARN · 0 stayed FAIL
  avg mids ΔE: 17.10 → 3.44 (80% reduction)
  avg max ΔE:  17.10 → 5.22

Round-1-sacred pattern:
  - Source PNG NEVER modified
  - Healed PNG → data/generated/midheal_<src>_<ts>.png
  - INSERT new all_designs row:
      generator='pil-mid-heal'
      parent_design_id=<src>
      is_published=TRUE
      kind='seamless_tile' (or whatever source had)
      notes='mid_heal of #<src>: PIL band-blur BAND=12 SIGMA=4.0'
  - Source notes appended: ' | MID_HEALED_BY: <new_id>'
  - Idempotent — skips already-healed sources via MID_HEALED_BY: tag

Reversible via POST /api/fix-decision/<new_id> verdict='reject'.

3-design smoke test passed before scaling: ids 24/4593/4594 healed,
new ids 43236/43237/43238 inserted with correct lineage, source notes
got MID_HEALED_BY append. Batch run on remaining 7,326 designs is in
flight at ~78 designs/s.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Files touched

Diff

commit 02f91e6f44d8923383a0a1210ef1829d16e77e04
Author: Steve Abrams <steve@designerwallcoverings.com>
Date:   Tue May 26 08:23:37 2026 -0700

    mid_seam_heal.py — $0 production fix for SDXL latent-grid leak
    
    Replaces the planned $293 Gemini-edit run on the 7,333 mids-only FAIL
    cohort. PIL Gaussian-band-blur (BAND=12px, SIGMA=4.0) at H/2 and V/2
    midlines collapses the 4x4 latent-block stitch discontinuity while
    preserving composition everywhere else.
    
    Validated on 20-sample experiment (committed e30db7df):
      20/20 improved · 9 → PASS · 11 → WARN · 0 stayed FAIL
      avg mids ΔE: 17.10 → 3.44 (80% reduction)
      avg max ΔE:  17.10 → 5.22
    
    Round-1-sacred pattern:
      - Source PNG NEVER modified
      - Healed PNG → data/generated/midheal_<src>_<ts>.png
      - INSERT new all_designs row:
          generator='pil-mid-heal'
          parent_design_id=<src>
          is_published=TRUE
          kind='seamless_tile' (or whatever source had)
          notes='mid_heal of #<src>: PIL band-blur BAND=12 SIGMA=4.0'
      - Source notes appended: ' | MID_HEALED_BY: <new_id>'
      - Idempotent — skips already-healed sources via MID_HEALED_BY: tag
    
    Reversible via POST /api/fix-decision/<new_id> verdict='reject'.
    
    3-design smoke test passed before scaling: ids 24/4593/4594 healed,
    new ids 43236/43237/43238 inserted with correct lineage, source notes
    got MID_HEALED_BY append. Batch run on remaining 7,326 designs is in
    flight at ~78 designs/s.
    
    Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 scripts/mid_seam_heal.py | 325 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 325 insertions(+)

diff --git a/scripts/mid_seam_heal.py b/scripts/mid_seam_heal.py
new file mode 100755
index 0000000..afbc48e
--- /dev/null
+++ b/scripts/mid_seam_heal.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+"""mid_seam_heal — production batch tool for the SDXL latent-grid leak.
+
+Defect: pixel discontinuity at row H/2 and col W/2 (the 4x4 latent-block
+stitch line). The 6-lens edges-agent flags this as `grid=mids_only` or
+`grid=both`. Average mids ΔE in flagged designs: 17.10.
+
+Fix: PIL Gaussian-blur of a 24-px band (12px each side) at H/2 and V/2.
+Composition preserved everywhere else. Validated on 20-design sample —
+20/20 improved, 0/20 stayed FAIL, 9/20 reached PASS, 11/20 reached WARN.
+Avg mids ΔE dropped 17.10 → 3.44 (80% reduction).
+
+Round-1-sacred pattern (per feedback_round_one_outputs_are_sacred):
+  - Source PNG NEVER modified or deleted
+  - Healed PNG written to data/generated/midheal_<src_id>_<ts>.png
+  - New spoon_all_designs row INSERTed:
+      generator='pil-mid-heal'
+      parent_design_id=<src_id>
+      is_published=TRUE
+      notes='mid_heal of #<src_id>: PIL band-blur BAND=12 SIGMA=4.0'
+  - Source kept is_published=FALSE (it already was, from the edges sweep)
+  - Source gets notes append: ' | MID_HEALED_BY: <new_id>'
+
+Reversible via POST /api/fix-decision/<new_id> with verdict='reject' —
+that flips new is_published=FALSE + restores source is_published=TRUE.
+
+Usage:
+  python3 mid_seam_heal.py --id 40689              # one design
+  python3 mid_seam_heal.py --all                   # every mids-only FAIL
+  python3 mid_seam_heal.py --all --workers 8       # parallel
+  python3 mid_seam_heal.py --grid both             # also heal 'both' cohort
+  python3 mid_seam_heal.py --limit 50              # cap for smoke test
+  python3 mid_seam_heal.py --dry-run               # plan only, no writes
+"""
+import argparse
+import json
+import multiprocessing as mp
+import os
+import subprocess
+import sys
+import tempfile
+import time
+from pathlib import Path
+
+WALLCO_ROOT = Path.home() / 'Projects' / 'wallco-ai'
+OUT_DIR = WALLCO_ROOT / 'data' / 'generated'
+BAND_PX = 12
+SIGMA = 4.0
+
+
+def fetch_targets(grid: str, limit: int | None, ids: list[int] | None):
+    """Return list of (id, local_path, dominant_hex, category, kind, palette_json)
+    rows for designs that match the mid-heal target.
+
+    grid=mids_only      → notes LIKE '%EDGES_FAIL: %grid=mids_only%'
+    grid=both           → both 'mids_only' AND 'both'
+    grid=any            → any EDGES_FAIL row (for diagnostic re-heal)
+    """
+    where = []
+    if ids:
+        where = [f"id IN ({','.join(str(int(i)) for i in ids)})"]
+    else:
+        # Skip rows that have already been mid-healed (avoid duplicate work)
+        where.append("(notes IS NULL OR notes NOT LIKE '%MID_HEALED_BY:%')")
+        # Skip user-removed
+        where.append("(user_removed IS NULL OR user_removed = FALSE)")
+        where.append("local_path IS NOT NULL")
+        if grid == 'mids_only':
+            where.append("notes LIKE '%EDGES_FAIL:%grid=mids_only%'")
+        elif grid == 'both':
+            where.append("(notes LIKE '%EDGES_FAIL:%grid=mids_only%' "
+                         "OR notes LIKE '%EDGES_FAIL:%grid=both%')")
+        elif grid == 'any':
+            where.append("notes LIKE '%EDGES_FAIL:%'")
+        else:
+            raise ValueError(f'bad --grid {grid}')
+
+    sql = (
+        "SELECT id, local_path, dominant_hex, category, kind, "
+        "COALESCE(palette::text, 'null') AS palette_json, "
+        "width_in, height_in, panels "
+        "FROM all_designs WHERE " + ' AND '.join(where)
+        + " ORDER BY id"
+    )
+    if limit:
+        sql += f' LIMIT {int(limit)}'
+    sql += ';'
+
+    r = subprocess.run(
+        ['psql', 'dw_unified', '-At', '-F', chr(31), '-c', sql],
+        check=True, capture_output=True, text=True,
+    )
+    rows = []
+    for line in r.stdout.splitlines():
+        if not line.strip():
+            continue
+        parts = line.split(chr(31))
+        if len(parts) < 9:
+            continue
+        try:
+            rows.append({
+                'id': int(parts[0]),
+                'local_path': parts[1],
+                'dominant_hex': parts[2] if parts[2] != '' else None,
+                'category': parts[3] or 'mixed',
+                'kind': parts[4] or 'seamless_tile',
+                'palette_json': parts[5] if parts[5] not in ('', 'null') else 'null',
+                'width_in': parts[6] if parts[6] != '' else None,
+                'height_in': parts[7] if parts[7] != '' else None,
+                'panels': parts[8] if parts[8] != '' else None,
+            })
+        except ValueError:
+            continue
+    return rows
+
+
+def heal_image(src_path: Path, out_path: Path):
+    from PIL import Image, ImageFilter
+    import numpy as np
+    img = Image.open(src_path).convert('RGB')
+    W, H = img.size
+    arr = np.asarray(img).copy()
+
+    blurred = img.filter(ImageFilter.GaussianBlur(radius=SIGMA))
+    blur_arr = np.asarray(blurred)
+
+    hm = H // 2
+    arr[hm - BAND_PX:hm + BAND_PX, :, :] = blur_arr[hm - BAND_PX:hm + BAND_PX, :, :]
+
+    wm = W // 2
+    arr[:, wm - BAND_PX:wm + BAND_PX, :] = blur_arr[:, wm - BAND_PX:wm + BAND_PX, :]
+
+    Image.fromarray(arr).save(out_path, 'PNG', optimize=True)
+    return out_path.stat().st_size
+
+
+def worker(row: dict):
+    """Heal one design. Returns dict with success metadata."""
+    src_id = row['id']
+    try:
+        src_path = Path(row['local_path'])
+        if not src_path.exists():
+            return {'id': src_id, 'ok': False, 'error': 'file_missing'}
+
+        ts = int(time.time() * 1000)
+        out_filename = f'midheal_{src_id}_{ts}.png'
+        out_path = OUT_DIR / out_filename
+
+        bytes_written = heal_image(src_path, out_path)
+
+        return {
+            'id': src_id,
+            'ok': True,
+            'out_path': str(out_path),
+            'out_filename': out_filename,
+            'bytes': bytes_written,
+            'category': row['category'],
+            'kind': row['kind'],
+            'dominant_hex': row['dominant_hex'],
+            'palette_json': row['palette_json'],
+            'width_in': row['width_in'],
+            'height_in': row['height_in'],
+            'panels': row['panels'],
+        }
+    except Exception as exc:
+        return {'id': src_id, 'ok': False, 'error': f'heal_exception: {exc}'}
+
+
+def sql_escape(s: str) -> str:
+    return s.replace("'", "''")
+
+
+def persist(results: list[dict]):
+    """Batched PG transaction:
+       INSERT new spoon_all_designs rows + UPDATE source notes for each success."""
+    ok = [r for r in results if r.get('ok')]
+    if not ok:
+        return 0
+
+    sql_parts = ['BEGIN;']
+    # Build the insert statements one at a time so we capture each new id.
+    # We use a CTE per row to also append to the source's notes column in
+    # the same batch. To keep this simple + atomic, we'll INSERT all rows
+    # via a TEMP TABLE join, then UPDATE sources.
+    sql_parts.append(
+        'CREATE TEMP TABLE _heal_in (src_id BIGINT PRIMARY KEY, '
+        'kind TEXT, category TEXT, dominant_hex TEXT, palette JSONB, '
+        'width_in INTEGER, height_in INTEGER, panels INTEGER, '
+        'local_path TEXT, image_url TEXT, prompt TEXT, seed BIGINT, '
+        'notes TEXT) ON COMMIT DROP;'
+    )
+
+    seed_base = int(time.time())
+    rows_for_sql = []
+    for i, r in enumerate(ok):
+        seed = (seed_base + i) & 0x7fffffff
+        kind = sql_escape(r['kind'])
+        cat = sql_escape(r['category'])
+        dh = "NULL" if not r['dominant_hex'] else f"'{sql_escape(r['dominant_hex'])}'"
+        pal_json = r.get('palette_json') or 'null'
+        if pal_json in ('', 'None'):
+            pal_json = 'null'
+        pal = f"'{sql_escape(pal_json)}'::jsonb"
+        wi = r['width_in'] if r['width_in'] not in (None, '') else 'NULL'
+        hi = r['height_in'] if r['height_in'] not in (None, '') else 'NULL'
+        pn = r['panels'] if r['panels'] not in (None, '') else 'NULL'
+        local_path = sql_escape(r['out_path'])
+        img_url = sql_escape(f"/designs/img/{r['out_filename']}")
+        prompt = sql_escape(f"mid_heal of #{r['id']}: PIL band-blur BAND={BAND_PX} SIGMA={SIGMA}")
+        notes = sql_escape(f"mid_heal of #{r['id']}: PIL band-blur BAND={BAND_PX} SIGMA={SIGMA}")
+        rows_for_sql.append(
+            f"({r['id']}, '{kind}', '{cat}', {dh}, {pal}, {wi}, {hi}, {pn}, "
+            f"'{local_path}', '{img_url}', '{prompt}', {seed}, '{notes}')"
+        )
+
+    CHUNK = 200
+    for i in range(0, len(rows_for_sql), CHUNK):
+        batch = rows_for_sql[i:i + CHUNK]
+        sql_parts.append(
+            'INSERT INTO _heal_in (src_id, kind, category, dominant_hex, palette, '
+            'width_in, height_in, panels, local_path, image_url, prompt, seed, notes) VALUES '
+            + ','.join(batch) + ';'
+        )
+
+    # Insert new rows + capture (src_id, new_id) for the source-notes update.
+    sql_parts.append("""
+WITH ins AS (
+  INSERT INTO all_designs
+    (kind, brand, width_in, height_in, panels, generator, prompt, seed,
+     image_url, local_path, dominant_hex, palette, category,
+     is_published, parent_design_id, notes)
+  SELECT
+    h.kind, 'wallco.ai', h.width_in, h.height_in, h.panels, 'pil-mid-heal',
+    h.prompt, h.seed, h.image_url, h.local_path, h.dominant_hex, h.palette,
+    h.category, TRUE, h.src_id, h.notes
+  FROM _heal_in h
+  RETURNING id AS new_id, parent_design_id AS src_id
+)
+UPDATE all_designs ad
+SET notes = CASE
+  WHEN ad.notes IS NULL OR ad.notes = '' THEN ' | MID_HEALED_BY: ' || ins.new_id::text
+  ELSE ad.notes || ' | MID_HEALED_BY: ' || ins.new_id::text
+END
+FROM ins
+WHERE ad.id = ins.src_id;
+""")
+
+    sql_parts.append('SELECT count(*) FROM _heal_in;')
+    sql_parts.append('COMMIT;')
+
+    full_sql = '\n'.join(sql_parts)
+    with tempfile.NamedTemporaryFile(mode='w', suffix='.sql', delete=False) as tf:
+        tf.write(full_sql)
+        sql_path = tf.name
+
+    r = subprocess.run(
+        ['psql', 'dw_unified', '-At', '-q', '-v', 'ON_ERROR_STOP=1', '-f', sql_path],
+        capture_output=True, text=True,
+    )
+    if r.returncode != 0:
+        print(f'PSQL FAILED:\n  stdout: {r.stdout[:500]}\n  stderr: {r.stderr[:500]}',
+              file=sys.stderr)
+        return 0
+    return len(ok)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--id', type=int, action='append', help='Specific id (repeatable)')
+    ap.add_argument('--all', action='store_true', help='Every matching design')
+    ap.add_argument('--grid', choices=['mids_only', 'both', 'any'], default='mids_only')
+    ap.add_argument('--limit', type=int, default=None)
+    ap.add_argument('--workers', type=int, default=8)
+    ap.add_argument('--dry-run', action='store_true', help='List targets, no writes')
+    args = ap.parse_args()
+
+    if not (args.id or args.all or args.limit):
+        ap.error('Pass --id, --all, or --limit')
+
+    targets = fetch_targets(args.grid, args.limit, args.id)
+    print(f'targets: {len(targets)} ({args.grid} cohort)', flush=True)
+    if not targets:
+        print('nothing to do', flush=True)
+        return
+
+    if args.dry_run:
+        print('--dry-run — first 10 ids:', [t['id'] for t in targets[:10]])
+        return
+
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+    t0 = time.time()
+    results = []
+    counts = {'ok': 0, 'err': 0}
+
+    with mp.Pool(args.workers) as pool:
+        for i, res in enumerate(pool.imap_unordered(worker, targets, chunksize=4), 1):
+            results.append(res)
+            if res.get('ok'):
+                counts['ok'] += 1
+            else:
+                counts['err'] += 1
+            if i % 250 == 0 or i == len(targets):
+                rate = i / max(0.001, time.time() - t0)
+                eta = (len(targets) - i) / max(0.001, rate)
+                print(
+                    f'  [{i:>5}/{len(targets)}] ok={counts["ok"]} err={counts["err"]} '
+                    f'rate={rate:.1f}/s eta={eta:.0f}s',
+                    flush=True,
+                )
+
+    pg_written = persist(results)
+    elapsed = time.time() - t0
+    print('', flush=True)
+    print(f'heal done in {elapsed:.1f}s — '
+          f'baked={counts["ok"]} err={counts["err"]} pg_inserted={pg_written}',
+          flush=True)
+
+    # Sample a few healed rows
+    sample = [r for r in results if r.get('ok')][:5]
+    for r in sample:
+        print(f"  id={r['id']:>6} → {r['out_filename']}  {r['bytes']//1024}KB", flush=True)
+
+
+if __name__ == '__main__':
+    main()

← 3190409 gitignore: exclude data/elements (6GB), data/rooms (141MB),  ·  back to Wallco Ai  ·  gitignore: exclude data/edges-scan-report-*.md (regeneratabl e69a345 →