Files
splatter-tennis/assets/_fine.py
T
wuwiandClaude Opus 5 0e92c72457 ONE palette gate replaces four feature-picked ones
Oliver, on the brute's fall frames: "now the fall zombie has a different skin
tone.. man this is all so tiresome." Both halves were right. The skin IS twenty
points lighter across that sheet, and the tiresome part is mine: I had been
adding one hand-picked gate per drift I discovered — redness, then hue, then
`changed`, then no-grids — which is a loop with no end, because naming features
is not the same as comparing pictures.

`kit` names four: shirt redness, shorts brightness, shorts hue, near-white
fraction. Every character breaks a different one. The brute's shirt has no red
in it at all, so its reference redness was 0 and blood scored 70-100 against a
tolerance meant for a garment. A body face down shows a fraction of the shirt it
shows standing, so two of the four move with POSE. And nothing in that list ever
measured SKIN, which is exactly how a brute's flesh drifted across a whole sheet
with six gates green.

The palette gate asks the only question that generalises: are the big colour
areas of this frame the same colours as the anchor's? Top SIX by area — the
garments and the skin. Proportions ignored, so a pose change is invisible; blood
is never in the top six, so a fresh wound cannot look like drift.

CALIBRATED BEFORE IT WAS WIRED IN, which is the discipline I skipped all day.
Frames known to be right score 0.0-8.5; the drifts we could actually see score
12.0-15.9. The threshold is 10, in the gap. Three formulations were measured and
the two that overlapped were thrown away.

And every sheet of a variant is now anchored to that variant's FIRST sheet, not
to the stage above it. Chaining let a legal hop compound into a drift nobody
authorised — the shorts walking -9 to -19 down the wound ladder was that, and so
was this.

Its first verdict on the existing brute: br_play is the character, and every
other sheet sits 15 to 48 away from it. It would have caught all of this on the
first roll.

Also in: the GRID guard from the same session — the valley-cut rescue would slice
a 2x2 sheet into four frames of a torso above someone else's boots, turning a
loud failure into a silent one. Asked per cut PIECE, because a body is vertically
continuous and a piece with a gap in it caught two of them.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-12 13:45:41 +02:00

195 lines
10 KiB
Python

# _fine.py — turn a generated sheet into FULL-RESOLUTION sprites.
#
# _slice.py resamples with NEAREST, which is right when the generator's art-pixel grid is coarse but
# shreds fine detail at a non-integer ratio. This one instead:
# 1. detects the art-pixel BLOCK SIZE the generator actually drew at, and snaps to that grid
# (one output pixel per art pixel — no half-blocks, no broken outlines),
# 2. common-scales the sheet to the target height with a box filter,
# 3. quantises every pixel back to the sheet's own palette, so the result is flat pixel art
# rather than a blurred photo of pixel art.
# Usage: _fine.py sheet.png out1.png out2.png ... [--h 224] [--cut X,Y,...]
import sys
from PIL import Image, ImageDraw
import numpy as np
TARGET_H = 224
KEY_D2 = 60 * 60
MIN_GAP = 12
MIN_BLOB = 60
def key(im, defringe=2):
a = np.array(im.convert('RGBA'))
bg = a[0, 0, :3].astype(int)
d = ((a[..., :3].astype(int) - bg) ** 2).sum(2)
magenta = np.minimum(a[..., 0], a[..., 2]).astype(int) > a[..., 1].astype(int) + 40
bgish = (d < KEY_D2) | magenta
# ONLY THE OUTSIDE IS BACKGROUND. This used to punch out every magenta pixel on the sheet
# wherever it was, and the generator paints a racket's STRINGS in the backdrop colour more often
# than not — so frame after frame installed a man holding an empty hoop. Background is the region
# CONNECTED TO THE BORDER (which touches_edge already guarantees no figure reaches), so an
# enclosed patch of it inside a racket head is not background and does not become a hole.
# .copy() is load-bearing: an Image.fromarray image is READ-ONLY, and ImageDraw.floodfill's
# first write raises inside its own try/except and returns having filled nothing — a silent
# no-op that looks exactly like "the whole sheet is one figure".
m = Image.fromarray(np.where(bgish, 255, 0).astype('uint8'), 'L').copy()
ImageDraw.floodfill(m, (0, 0), 128)
outside = np.array(m) == 128
if not outside.any(): sys.exit('ERROR: background flood filled nothing — is (0,0) not background?')
a[outside] = 0
# ...and what that leaves — magenta INSIDE a body — is repainted as pale grey rather than kept,
# because bright pink strings are as wrong as no strings. Luminance is preserved, so the drawn
# crosshatch survives as crosshatch; only the hue goes.
inner = magenta & ~outside
if inner.any():
lum = a[..., :3].mean(2)
a[inner, 0] = a[inner, 1] = a[inner, 2] = np.clip(lum[inner] * 0.5 + 115, 0, 255).astype('uint8')
# De-fringe. The generator does not always give you a pure key colour, and whatever it does give
# bleeds a halo one pixel into the sprite — which survives as a magenta rim once the sprite is on
# a green court. Anything on the EDGE that is still leaning towards the background goes too.
# ...but each pass also eats a 1-pixel line, and a racket frame IS a 1-pixel line: two passes
# removed the rackets from a whole serve sheet and left men holding a grip. --defringe 1 (or 0)
# is for sheets whose art is that thin.
for _ in range(defringe):
solid = a[..., 3] > 0
edge = solid & ~(np.roll(solid, 1, 0) & np.roll(solid, -1, 0) & np.roll(solid, 1, 1) & np.roll(solid, -1, 1))
near = ((a[..., :3].astype(int) - bg) ** 2).sum(2) < KEY_D2 * 9
a[edge & near] = 0
return a
def block_size(a):
ink = a[..., 3] > 0
ys, xs = np.where(ink)
sub = a[ys.min():ys.max()+1, xs.min():xs.max()+1]
h, w = sub.shape[:2]
best = 1
for b in range(2, 13):
if h < b * 6: break
ok = tot = 0
for y in range(0, h - b, b):
for x in range(0, w - b, b):
blk = sub[y:y+b, x:x+b].reshape(-1, 4)
tot += 1
if (blk == blk[0]).all(): ok += 1
if tot and ok / tot > 0.82: best = b
return best
# WHERE TO CUT WHEN THE FIGURES TOUCH. Segmentation by empty columns needs a clear gap, and the
# generator routinely spaces a row so a belly, an elbow or a racket overlaps its neighbour — a whole
# sheet of otherwise perfect art then reports "found 2 frames, expected 4" and is thrown away. That
# happened five times in one afternoon, and the sheet it lost last was the best roll of the day.
# A row is EVENLY SPACED by construction (we ask for one), so when the count is known the honest
# recovery is: look near each even division and cut at the THINNEST column within that window —
# between two figures the ink thins out even where it does not vanish. Never used to guess a count,
# only to reach a count we already asked for.
def valley_cuts(ink_cols, x0, x1, n):
span = x1 - x0
win = max(2, span // (3 * n))
out = []
for k in range(1, n):
c = x0 + span * k // n
lo, hi = max(x0 + 1, c - win), min(x1 - 1, c + win)
if hi <= lo: return []
out.append(lo + int(np.argmin(ink_cols[lo:hi])))
return out
# DO THE FIGURES SHARE A GROUND LINE? That is the invariant every prompt in this project asks for,
# and it is the one that tells a ROW from a GRID. An empty-band test does not: in the sheet that
# caught this, a 2x2 of a brute falling, the bottom-right figure was tall enough to overlap the top
# row and there was no clean horizontal gap anywhere.
#
# It matters because the valley rescue below will cut ANY sheet into the number of columns it was
# asked for — and on a grid each column is a torso above someone else's boots. Four such frames
# installed silently and looked, in the log, exactly like a success. The rescue turned a loud failure
# into a quiet one, which is the worst trade available in this file; this is the check that stops it.
# IS THIS ONE FIGURE, OR TWO STACKED? Asked of each CUT PIECE, not of the sheet. A body is
# vertically continuous — head to feet there is no height at which the piece is empty — so a
# full-width empty band inside a piece's own extent means the cut caught someone's torso above
# someone else's boots.
#
# Asking it of the whole sheet does not work, and both weaker versions were tried: an empty-band test
# on the sheet fails when a grid's bottom-right figure is tall enough to overlap the top row, and a
# shared-ground-line test fails because after a bad cut EVERY column reaches the bottom row's feet.
# The piece is the only place the question has a clean answer.
def split_piece(crop):
rows = (crop[..., 3] > 0).any(1)
ys = np.where(rows)[0]
if not len(ys): return False
run = best = 0
for y in range(ys.min(), ys.max() + 1):
run = 0 if rows[y] else run + 1
if run > best: best = run
return best >= MIN_GAP * 2
def frames(a, cuts, want=0):
ink = (a[..., 3] > 0)
cols = ink.any(0)
out, start, gap = [], None, 0
for x in range(len(cols)):
if cols[x]:
if start is None: start = x
gap = 0
elif start is not None:
gap += 1
if gap >= MIN_GAP: out.append((start, x - gap + 1)); start = None
if start is not None: out.append((start, len(cols)))
for c in sorted(cuts, reverse=True): # manual splits where two figures touch
for i, (x0, x1) in enumerate(out):
if x0 < c < x1: out[i:i+1] = [(x0, c), (c, x1)]; break
# ...and the automatic version of the same rescue, when no manual cut was given
if want and len(out) < want and not cuts:
xs = np.where(cols)[0]
if len(xs):
auto = valley_cuts(ink.sum(0), int(xs.min()), int(xs.max()) + 1, want)
if auto:
print(f'note: only {len(out)} blob(s) — figures touch; cutting at {auto}')
edges = [int(xs.min())] + auto + [int(xs.max()) + 1]
out = list(zip(edges, edges[1:]))
return out
def main(sheet, outs, target_h, cuts, anchor, defringe=2):
a = key(Image.open(sheet), defringe)
b = block_size(a)
crops = []
spans = frames(a, cuts, len(outs))
for x0, x1 in spans:
sub = a[:, x0:x1]
ys = np.where((sub[..., 3] > 0).any(1))[0]; xs = np.where((sub[..., 3] > 0).any(0))[0]
crops.append(sub[ys.min():ys.max()+1, xs.min():xs.max()+1])
if len(crops) != len(outs):
sys.exit(f'ERROR: found {len(crops)} frames, expected {len(outs)} ({sheet})')
bad = [i for i, c in enumerate(crops) if split_piece(c)]
if bad:
sys.exit(f'ERROR: frame(s) {bad} contain two stacked bodies — the sheet is a GRID, not one '
f'row ({sheet})')
# palette = the colours the generator actually used, on the largest frame
big = max(crops, key=lambda c: (c[..., 3] > 0).sum())
px = big[big[..., 3] > 200][:, :3]
pal = np.unique((px // 8) * 8, axis=0)
# ONE scale for the sheet, anchored on a chosen frame — the STANDING pose is the 224 reference,
# so a smash with the arm overhead stays legitimately taller (the game scales off img.height).
anchor_h = crops[anchor].shape[0] if 0 <= anchor < len(crops) else max(c.shape[0] for c in crops)
k = target_h / (anchor_h / b)
for c, out in zip(crops, outs):
h, w = c.shape[:2]
snap = Image.fromarray(c, 'RGBA').resize((max(1, round(w / b)), max(1, round(h / b))), Image.BOX)
fh = max(1, round(snap.height * k)); fw = max(1, round(snap.width * k))
im = snap.resize((fw, fh), Image.BOX)
arr = np.array(im).astype(int)
solid = arr[..., 3] > 110
flat = arr[..., :3].reshape(-1, 3)
idx = ((flat[:, None, :] - pal[None, :, :]) ** 2).sum(2).argmin(1)
arr[..., :3] = pal[idx].reshape(arr.shape[0], arr.shape[1], 3)
arr[..., 3] = np.where(solid, 255, 0)
Image.fromarray(arr.astype('uint8'), 'RGBA').save(out)
print(f'{out}: {fw}x{fh} (art block {b} in source)')
if __name__ == '__main__':
args = sys.argv[1:]
th, cuts, anchor = TARGET_H, [], -1
if '--h' in args: i = args.index('--h'); th = int(args[i+1]); del args[i:i+2]
if '--cut' in args: i = args.index('--cut'); cuts = [int(v) for v in args[i+1].split(',')]; del args[i:i+2]
de = 2
if '--defringe' in args: i = args.index('--defringe'); de = int(args[i+1]); del args[i:i+2]
if '--anchor' in args: i = args.index('--anchor'); anchor = int(args[i+1]); del args[i:i+2]
main(args[0], args[1:], th, cuts, anchor, de)