Six sessions of asking an image model for a row of running figures, and it
never once drew a gait: three phases repeated with the same leg leading,
"ropeskipping, not running". A frame-by-frame LEFT/RIGHT skeleton spec, a
closing checklist, a loop clause, six figures instead of eight — all lost.
One 5-second image-to-video clip has no such problem, because it moves a
body instead of drawing poses. ~$0.42 a character through the key already in
.env, pose pinned at BOTH ends (pinned only at the front, she rotated into
profile halfway), stride period MEASURED off the frames rather than sampled
by eye — hers repeats in 6 clip frames, his in 23 — and the phases cut and
palette-snapped through _stride.cut. assets/_vidroll.py films and cuts it;
her first frame came from one still roll that took his approved pose as a
picture instead of a paragraph.
Wired: runN probes the sheet through SK() so the women's tour counts HER
frames, the player is no longer clamped to a two-frame pair, runHz 2.4 ->
1.9 ("this looks like roadrunner"), and the bench grew two PLAYER stops —
a cycle only ever seen mid-rally is a cycle nobody has watched.
Also in: the difficulty retune (medium is a real step above the shipped CPU
and is where `smart` first appears; hard stepped back off superhuman; easy
closed its gap) with MEDIUM's dial-for-dial pin to TUNE removed from the
test, and Oliver's own transparency fixes on the male frames.
SPRV 138 — the frames were right on disk and wrong on screen for an hour
because assets are cached for a year unless the ?v= moves. main.js v269.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Oliver, after the facing fix: still flashing, female worst. Film-strip
diagnosis in the pane: with facing pinned the racket STILL alternated —
frame 4 of the female cycle tucked it across the body, and half the frames
of both player cycles lost or shrank the hoop to cut fragments (overlapping
figures sliced by the even-spaced valley cuts).
Both cycles re-rolled with three new prompt rules that now live in
_runroll.py: the racket stays right of the midline in EVERY figure (named
again for the airborne pose, where it kept vanishing), and the figures
never overlap (no more sliced hoops — a debris-stripping pass was tried
and removed: after the recrop a real racket is indistinguishable from a
fragment, prevention beats cleanup). The model returns 2-row grids at
times; a coherent band is cropped and cut. Installed: racket-side spread
+0.01..+0.05 across all frames of both cycles (was a full side swap).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver: 'the motion seems very flashy and unnatural'. Two causes, both
measured: drawChar centres a frame by its total width while the racket's
reach varies across the cycle, so the BODY teleported sideways up to 28
source px on a frame swap (the female player was worst) — every frame is now
padded until its lower-third centroid sits at the frame centre (residuals
±0.2 px, the normalisation lives in _runroll.py's cut). And the hardcoded
4.5 cycles/s was tuned on 2-4 near-duplicate frames; on eight real phases it
was a strobe — rolled sets stride at TUNE.runHz (2.4), old art keeps 4.5.
In-pane: body wobble down from ~36 device px per swap to 3.2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The old run frames were generated one at a time and it showed: 91/110/92/98 px
of body-width drift, duplicated poses drawn differently, no racket. Each
character (player m/f 6 frames, opponent m/f 8) now comes from ONE 4K call
anchored on a composed identity reference of its settled frames — the broken
run art stayed out of the reference on purpose. Stride-gated, cut with _fine,
judged by eye per roll. drawChar probes run8/run6/run4/run2.
Side effect: the male opponent's old run art faced LEFT against the
art-faces-RIGHT convention — he had been moonwalking right since install.
Fixed by drawing the new rows to the convention.
Pipeline lessons recorded in PROJECT.md (the model negotiates the count; a
back view has no direction; the racket clause goes early and loud).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The playtest list, art side. The generator had been copying the very frames I
wanted changed, so:
- The player is drawn FROM BEHIND again in every pose — standing, bouncing,
serving, swinging. He had turned to face the camera in the redraw, which is
also why he "looked at you" in different poses.
- The serve frames are back at full size. That batch had been anchored on the
toss frame, so the whole serve set came out about a fifth too small.
- The runner CHARGES AT YOU down the court instead of sprinting sideways across
it, which is what he actually does in the game.
- The brute's wound stages are drawn a pair at a time from their own A frame, so
the injury stops moving between the two feet.
- The bounce frames had a ball painted into them while the game draws a live one
on top — he was bouncing two. The painted one is gone, and the live one now
falls on the side the sprite actually reaches for.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both full sets through the same pipeline: stances and ball-bounce, serve toss,
smash and both swing frames, the four-frame run, and the five-frame death. The
opponent's ten frames with them. Inner-edge density roughly doubles across the
board — the serve frames go 0.22-0.32 to 0.41-0.54.
_regen.py retries a rejected batch itself now (default three attempts). A
rejection is usually just a bad roll — two figures drawn close enough that the
slicer sees one frame, or a limb off the canvas — and re-rolling fixes it without
me sitting in the loop.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two frames is a lurch, and he is on screen the whole run — during a horde he is
the only thing moving. Regenerated the whole cycle in one pass rather than
bolting two new frames onto the old pair, so the four are consistent with each
other, then fitted them per channel onto the mean and sd of the frames that
stay (idle + swing): mean error under 3 of 255, so he does not change colour
when he starts moving.
The pick is frame-count driven, not hardcoded — the opponent has no run3/run4
and stays on two, and the cycle RATE is unchanged either way: more frames
subdivide the stride, they do not speed it up.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The umpire is situational again. Each line is tied to the thing that justifies
it — a new variant appearing, the court filling, one crossing your baseline, a
game lost, a horde, a changeover — behind one shared cooldown, so several
triggers at once still yield one remark. The random timer pool is gone, the
interval went 26s to 55s, and ump_out_of_order, ump_let and ump_no_appeal are
deleted.
"Visually the ball is not yet at the player but somehow the player already is
hit" was real and measurable. The hostile-ball hit test reused the parry's
hypot(dx, dz*0.8), and weighting depth shrinks it: contact fired while the ball
was still 0.68 m up-court and drawn up around his head. It has its own
unweighted test now, and contact lands at 0.25 m with the ball on his chest.
The parry keeps its weighting — a racket sweep really is wider than it is deep.
On "everything is a tad too small": measured first, because higher-resolution
sprites are only half the answer. At RES 3 the player rasterises into 170
device pixels from a 112px source — upscaled 1.5x — while a far shambler is
downscaled to 78. So near bodies needed resolution and far bodies needed the
camera.
Camera: pushed back and lengthened (camZ 42->48, camF 580->760) to flatten the
perspective so distant bodies grow relative to near ones rather than everything
zooming, camY raised for a steeper look-down, horizon re-centred. Far shambler
26->30px, boss 39->45px, player unchanged, court in the same band.
Sprites: the original sheets were still in the repo at full resolution, so
nothing needed regenerating — IDLE_TARGET_H went to 224 and every sheet was
re-sliced. The sheet-to-output mapping was never recorded, so it was recovered
by matching sliced frames to existing assets by image signature: 58 came back
from sheets, the 29 without a confident match were 2x-upscaled to stay on the
shared scale, since a sprite off that scale renders at half height.
Tennis untouched: 50 hits / 0 whiffs. 45-game soak clean, all four chapters.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- white ring shows where the incoming ball will land; yellow/red ring shows
where your shot will land (red = going out) — Mario-Tennis-style feedback
- ball boys wait at the net posts, fetch every dead ball, and trot home
- canvas renders at 2x density, all sprites re-sliced at 112px; run/swing
sheets regenerated in the finer style
- assist: 0.35m deadzone + facing hysteresis kills the left/right flicker;
player reach 1.5 -> 1.7
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
wuwiandClaude Opus 5
