RETARGET-AB.md
Source game/motion/RETARGET-AB.md. Last git commit 2026-11-12. About 2,382 words.
Retarget A/B: UE IK Retargeter vs toyxyz-blender-retargeter
Run 2026-11-12. Measure only, nothing imported into the game project. Scripts: ab_ue_export.py, ab_ue_incumbent.py, ab_challenger.py, ab_measure.py, ab_report.py, ab_strip.py (this folder). Raw data and strips: D:\megajam\retarget\ab\.
Setup
- Takes (Kimodo SOMA77, tower 1): take_chase_split_a-s79, take_chase_split_c-s36, take_cannon_full-s5, brawl_beat_b-s1. Source npz and bvh copied to
ab\in. - Incumbent: Blender BVH to FBX (
retarget_blender_fbx.py, yaw 0), UE IK RetargeterRTG_SOMA_to_MH, commit 7f7014a chain. Three takes were already at /Game/Motion/Kimodo and were read back as they are. brawl_beat_b-s1 was imported and retargeted with the same retargeter into a throwaway /Game/ABTrial folder, read back, and the folder was deleted. The read-back isretarget_export.pyunchanged (rig-export npz). - Challenger: toyxyz-blender-retargeter 0.4.30 (GPL-3.0, cloned to
tools-ext/toyxyz-retargeter, git-ignored, run as a tool only). Blender 5.2.2 headless. Target is the MetaHuman body skeleton exported from SKM_Body byab_ue_export.py. Solver settings are the add-on defaults (strength 100, root travel Scale to Proportions, floor contact on, floor World Z = 0). Bone mapping is the add-on's own Build Bones List (it matched SOMA and MetaHuman by name, 3 source spine bones to 5 target ones, no manual edits). Baked per frame to an Action, then read back in the same rig-export layout. - Two reference-pose steps were needed and are scripted in
ab_challenger.py: (1) the BVH rest has the hips at the origin and the legs below the floor, so the source reference is the rest rotations with the hips lifted until the lowest bone touches the floor (without it the target floats 0.9 m up); (2) the MetaHuman rest is an A-pose, so the target reference is a T-pose made by raising both upper arms to horizontal and straightening both forearms. Trial of the three target references on brawl_beat_b-s1 / take_chase_split_c-s36, joint-limit gate violations: A-pose as imported 612 / 609; upper arms raised only 188 / 66; upper arms raised and forearms straight 0 / 6 (used). That choice was made by looking at the gate on those two takes, so the challenger got a tuned setup. The incumbent was not tuned in this trial. - Measured on the same code for both: both results are rig-export npz files (joint_limits.py layout, metres, z up, MetaHuman bones only: pelvis, three spine, neck, head, clavicles, arms, hands, middle metacarpal and finger 1, legs, feet, balls; no other fingers), cut to the same frame count (the UE importer drops the last frame, 149 vs 150). Comparing as npz rather than as Blender actions removes a coordinate conversion; the incumbent never went through Blender as an action.
- Foot slide: per bone (foot and ball, both sides), frames within 3 cm of that bone's minimum height in the take and under 0.2 m/s; contiguous runs of 2 or more frames are one contact; slide is the horizontal path length inside a contact, in cm. The "height test only" row drops the speed test, so it also counts a foot that skates quickly while low.
- Root drift: pelvis horizontal path, in each take's own start frame, against the source path scaled by 0.858 (target standing hip height 0.871 m over source 1.015 m). Path ratio is path length over scaled source path length.
- Gate:
joint_limits.py --no-fingerson each npz. Wrist twist is the gate's ownwrist_*.twistmeasure (relative to each rig's rest), so the min and max carry each rig's rest offset; compare the range.
Results
take_chase_split_a_s79 (149 frames at 30 fps)
| Measure | Kimodo source | Incumbent | Challenger |
|---|---|---|---|
| Foot slide, cm per contact, foot_l | 0.59 (n=6, max 3.2) | 1.32 (n=5, max 4.4) | 0.92 (n=5, max 3.1) |
| Foot slide, cm per contact, ball_l | 1.59 (n=7, max 3.3) | 1.44 (n=6, max 4.0) | 1.43 (n=7, max 2.9) |
| Foot slide, cm per contact, foot_r | 0.40 (n=4, max 1.2) | 1.67 (n=4, max 3.0) | 0.67 (n=4, max 1.2) |
| Foot slide, cm per contact, ball_r | 1.46 (n=4, max 2.1) | 1.75 (n=5, max 2.9) | 1.31 (n=5, max 2.9) |
| Foot slide, cm per contact, all | 1.05 (n=21, max 3.3) | 1.53 (n=20, max 4.4) | 1.14 (n=21, max 3.1) |
| Foot slide, height test only (speed test dropped), cm per contact, all | 10.6 (n=29) | 6.0 (n=29) | 13.5 (n=31) |
| Root path length / (source path x 0.8578) | 1.000 (source 10.393 m) | 1.12 | 1.062 |
| Root drift, end error vs scaled source path, cm | - | 105.3 | 53.7 |
| Root drift, max / rms deviation, cm | - | 105.3 / 69.9 | 53.7 / 35.6 |
| Joint-limit gate, joint-frame violations (frames affected) | 0 (0 of 149) | 0 (0 of 149) | 3 (3 of 149) |
| Gate, joints over | none | none | shoulder_l.behind x2, shoulder_r.behind x1 |
| Wrist twist, left: range deg (min to max) | 56.5 (-37.0 to 19.5) | 56.5 (-34.5 to 22.0) | 56.7 (-19.3 to 37.4) |
| Wrist twist, right: range deg (min to max) | 42.6 (-15.2 to 27.4) | 41.8 (-41.5 to 0.3) | 43.0 (-27.0 to 16.0) |
take_chase_split_c_s36 (149 frames at 30 fps)
| Measure | Kimodo source | Incumbent | Challenger |
|---|---|---|---|
| Foot slide, cm per contact, foot_l | 0.20 (n=5, max 0.6) | 1.44 (n=5, max 3.2) | 0.38 (n=5, max 0.9) |
| Foot slide, cm per contact, ball_l | 1.88 (n=6, max 4.4) | 2.48 (n=5, max 4.2) | 1.56 (n=6, max 3.4) |
| Foot slide, cm per contact, foot_r | 0.60 (n=5, max 1.9) | 1.21 (n=5, max 2.8) | 0.57 (n=6, max 1.9) |
| Foot slide, cm per contact, ball_r | 0.83 (n=7, max 1.6) | 1.19 (n=6, max 2.4) | 0.74 (n=8, max 1.3) |
| Foot slide, cm per contact, all | 0.92 (n=23, max 4.4) | 1.56 (n=21, max 4.2) | 0.82 (n=25, max 3.4) |
| Foot slide, height test only (speed test dropped), cm per contact, all | 7.2 (n=22) | 6.3 (n=23) | 8.2 (n=21) |
| Root path length / (source path x 0.8578) | 1.000 (source 5.222 m) | 1.12 | 1.058 |
| Root drift, end error vs scaled source path, cm | - | 50.6 | 26.1 |
| Root drift, max / rms deviation, cm | - | 51.0 / 40.9 | 27.0 / 21.3 |
| Joint-limit gate, joint-frame violations (frames affected) | 2 (2 of 149) | 8 (6 of 149) | 6 (5 of 149) |
| Gate, joints over | wrist_l.twist x2 | wrist_l.side x2, wrist_l.twist x4, ankle_l.twist x2 | wrist_l.side x2, wrist_l.twist x2, ankle_l.twist x2 |
| Wrist twist, left: range deg (min to max) | 116.3 (-35.9 to 80.4) | 115.2 (-94.8 to 20.4) | 116.0 (-80.0 to 36.0) |
| Wrist twist, right: range deg (min to max) | 88.0 (-60.5 to 27.5) | 88.1 (-42.9 to 45.2) | 88.3 (-27.8 to 60.5) |
take_cannon_full_s5 (270 frames at 30 fps)
| Measure | Kimodo source | Incumbent | Challenger |
|---|---|---|---|
| Foot slide, cm per contact, foot_l | 1.02 (n=4, max 3.0) | 2.86 (n=5, max 4.8) | 1.91 (n=4, max 3.2) |
| Foot slide, cm per contact, ball_l | 3.44 (n=4, max 6.2) | 5.44 (n=4, max 8.2) | 3.43 (n=4, max 4.8) |
| Foot slide, cm per contact, foot_r | 0.84 (n=3, max 1.5) | 2.87 (n=3, max 3.0) | 0.92 (n=4, max 1.9) |
| Foot slide, cm per contact, ball_r | 2.12 (n=7, max 5.7) | 3.71 (n=4, max 6.5) | 2.92 (n=5, max 5.5) |
| Foot slide, cm per contact, all | 1.96 (n=18, max 6.2) | 3.72 (n=16, max 8.2) | 2.33 (n=17, max 5.5) |
| Foot slide, height test only (speed test dropped), cm per contact, all | 10.7 (n=15) | 11.2 (n=15) | 13.9 (n=12) |
| Root path length / (source path x 0.8578) | 1.000 (source 1.735 m) | 1.015 | 1.087 |
| Root drift, end error vs scaled source path, cm | - | 0.2 | 0.8 |
| Root drift, max / rms deviation, cm | - | 0.9 / 0.5 | 4.2 / 2.1 |
| Joint-limit gate, joint-frame violations (frames affected) | 0 (0 of 270) | 0 (0 of 270) | 0 (0 of 270) |
| Gate, joints over | none | none | none |
| Wrist twist, left: range deg (min to max) | 58.4 (-35.9 to 22.5) | 58.2 (-38.2 to 20.0) | 58.7 (-22.3 to 36.4) |
| Wrist twist, right: range deg (min to max) | 26.9 (-16.3 to 10.6) | 26.5 (-25.9 to 0.6) | 27.4 (-10.3 to 17.1) |
brawl_beat_b_s1 (149 frames at 30 fps)
| Measure | Kimodo source | Incumbent | Challenger |
|---|---|---|---|
| Foot slide, cm per contact, foot_l | 0.49 (n=2, max 1.0) | 3.05 (n=2, max 4.2) | 1.10 (n=2, max 1.9) |
| Foot slide, cm per contact, ball_l | 3.20 (n=2, max 4.5) | 3.97 (n=2, max 5.4) | 2.71 (n=2, max 3.8) |
| Foot slide, cm per contact, foot_r | 0.55 (n=3, max 1.2) | 1.87 (n=3, max 3.6) | 0.69 (n=3, max 1.0) |
| Foot slide, cm per contact, ball_r | 1.94 (n=3, max 2.5) | 2.56 (n=3, max 4.7) | 1.94 (n=3, max 2.9) |
| Foot slide, cm per contact, all | 1.48 (n=10, max 4.5) | 2.73 (n=10, max 5.4) | 1.55 (n=10, max 3.8) |
| Foot slide, height test only (speed test dropped), cm per contact, all | 19.2 (n=14) | 25.7 (n=12) | 24.6 (n=13) |
| Root path length / (source path x 0.8578) | 1.000 (source 1.826 m) | 1.015 | 1.059 |
| Root drift, end error vs scaled source path, cm | - | 2.1 | 8.3 |
| Root drift, max / rms deviation, cm | - | 2.1 / 1.7 | 8.4 / 6.9 |
| Joint-limit gate, joint-frame violations (frames affected) | 0 (0 of 149) | 0 (0 of 149) | 0 (0 of 149) |
| Gate, joints over | none | none | none |
| Wrist twist, left: range deg (min to max) | 13.2 (21.5 to 34.7) | 13.1 (-49.9 to -36.8) | 13.4 (-34.4 to -21.0) |
| Wrist twist, right: range deg (min to max) | 17.2 (11.1 to 28.3) | 17.4 (-43.9 to -26.5) | 17.4 (-28.2 to -10.8) |
Summary
| Measure | Incumbent | Challenger | Reads as |
|---|---|---|---|
| Foot slide, mean cm per contact (gated), chase a / chase c / cannon / brawl | 1.53 / 1.56 / 3.72 / 2.73 | 1.14 / 0.82 / 2.33 / 1.55 | challenger lower on 4 of 4 takes and on all 16 bone cells; source itself is 1.05 / 0.92 / 1.96 / 1.48 |
| Foot slide, height test only | 6.1 / 6.3 / 11.3 / 25.7 | 13.5 / 8.2 / 13.9 / 24.6 | incumbent lower on 3 of 4; the speed-free count does not back the win |
| Root end error, cm (chase a / chase c / cannon / brawl) | 105 / 51 / 0.2 / 2.1 | 54 / 26 / 0.8 / 8.3 | challenger closer on the two long runs, incumbent closer on the two short takes |
| Gate violations (joint-frames), source in brackets | 0 (0) / 8 (2) / 0 (0) / 0 (0) | 3 (0) / 6 (2) / 0 (0) / 0 (0) | tie overall (9 vs 8); challenger needs the straight-forearm T-pose reference |
| Wrist twist range vs source, both wrists, all four takes | within 1.1 degrees | within 0.5 degrees | tie |
Verdict
No swap. The challenger wins one measure and ties the rest.
- Foot slide: the challenger is lower on all four takes, by 0.4 to 1.4 cm per contact (1.14 vs 1.53, 0.82 vs 1.56, 2.33 vs 3.72, 1.55 vs 2.73), and sits near the Kimodo source's own slide. That is a real but small difference: the incumbent's worst takes slide 2.7 to 3.7 cm per contact, which is hard to see at game scale. The speed-free check does not repeat the win (incumbent lower on three of four), so the claim is limited to the slow creep the brief defines.
- Root drift: split. The challenger travels 6 to 9 percent past the scaled source on every take. The incumbent travels 12 percent past on the two chase runs and 1.5 percent on the other two, so it is closer on the short takes and further off on the long ones. Neither is a win.
- Joint-limit gate: tie, 9 versus 8 joint-frames in total across four takes of about 700 frames. The challenger only passes with the hand-made reference pose; with the A-pose as imported it fails 600 joint-frames per take (elbows folded or twisted).
- Wrist twist: both follow the source range on all eight wrist ranges (two wrists, four takes), the incumbent within 1.1 degrees and the challenger within 0.5. Their min and max sit at different offsets from the source because the rest hand orientations differ, which is why the table compares ranges.
- Stick figures (
D:\megajam\retarget\ab\strip_*.png) show the two results on top of each other everywhere except the root offset in the long takes. I looked at brawl, cannon and chase_c (chase_a not opened); no visibly broken pose in the final challenger run (the first two reference setups did fold the right forearm in brawl, which is what the gate caught).
Cost of switching: a Blender step with a manual reference pose per skeleton, an FBX export of the baked Action back to Unreal, and a second place to keep in sync with the MetaHuman skeleton. For a 0.4 to 1.4 cm foot-slide gain that does not pay. If foot slide on a specific shot is the problem, run this add-on on that take only. Not tested here: finger and toe retargeting, FBX round trip into UE, takes with the hands planted (the add-on lists hand contact as a known gap), and the add-on at non-default solver settings.
README.md
Source game/sets/README.md. Last git commit 2026-11-13. About 2,077 words.
Sets (CONTENT-PLAN section 1)
Four dressed areas of Stage 7, each placed from a layout file by one script: stage7 (the working stage between takes), midway (chase), beertent (brawl), cannonfield (cannon). Built into the game since the set build ticket (2026-11-13).
Where the action is (45 s takes, commit 3dfe3f1)
The kid's paths come from packaged perfect-bot runs of full5 with -KKPelvisLog (D:\megajam\sets_work\pelvis\), with the samples that stray more than 120 cm from the director build's own keyed path left out (PLAN_TOL): in full5 two takes played a baked clip of another take (see "Shared clip names" below), which put the chase kid at the origin from 33.7 to 38.7 s.
Nothing below 300 cm stands in these boxes (world cm, L_Stage7):
| take | x from | x to | y from | y to | what is there |
|---|---|---|---|---|---|
| chase | 2355 | 3555 | -375 | 342 | the mirror maze run and the tailgate leap onto the flatbed (the first layout's box; the new action stayed in it) |
| brawl | from the path | the kid's and the heavy's reach from 21 s plus the corridor, at least x -290 to 260, y -200 to 150: the rope swing, the uppercut and the bandstand fall | |||
| cannon | -1411 | -211 | -196 | 204 | the run-in behind the cannon and the crowd wave (the first layout's box) |
The bandstand fall lands the kid near (-16, -44) at 31 s. The bandstand is placed by place_fall_target: as close to the landing as the rules allow, the fall itself exempt. The closest legal spot is (-136, 216), its front edge 110 cm behind the landing: the chair section (14 to 22 s) runs the kid through x -150 to -184 right where the bandstand would have to stand to catch him. To land the fall on it, the fall has to carry him about a metre further back (north), or the chair section has to move.
Take props (proposed/take_props.diff, applied)
- chase: popcorn cart (squashed to 62 %: the vault), swing gate (turned and stretched to the slide's 220 x 110), the batch 15 flatbed as the truck. The flatbed's bed is at 114 cm (measured from the GLB), not the box truck's 160, so the kid's three "lift 160" sections stand him at 114; the truck is staged 50 cm further ahead both times so its tailgate (650 long against the box's 550) is where the leaps were timed.
- brawl: the rope is a real mesh (
meshgen/make_rope.py: hemp, three-strand twist painted in its texture, an overhand knot above a frayed end), hanging 200 to 460 cm like the stand-in. - cannon: the batch 15 platform with its deck raised from 117 to 200 cm (
meshgen/raise_platform.py: only the legs, braces and ladder get longer; deck and rails as made), turned so its ladder faces -Y where the kid climbs (16.5 to 17.5 s, from (20, -104) up to (0, 0)), 0.8 along X so the deck keeps clear of the cannon like the stand-in. One whole flatbed replaces the three truck boxes (bed 2075 to 2430, under the landing at 2258 to 2344), the crash mats on its bed at x 2290. frame_checks point at it.
Run order (laptop, no GUI)
The game patch (proposed/KKGame_settags.patch, applied) drops the other takes' sets by their KK_settake_<take> tag, so set actors no longer carry DIR_<take> and a director rebuild leaves them alone. A walk-through take (walk_<take>) keeps its take's set.
python layout_tool.py all(only after a take changes): layouts, plots, the take diff, the check.run_ue_py.ps1 -Script C:\Users\owner\NEWBOY\lab\megajam\game\sets\set_build.py -ScriptArgs "--import-only"imports every GLB inmeshes.jsoninto/Game/Sets/<home set>/, Nanite on for meshes of 10k triangles and up (TRELLIS props), simple collision removed (the actors are NO_COLLISION too). Thenset_slim.pywith-ScriptArgs "apply"(the size budget below).- Director builds, as usual (
run_director_build.ps1 -Take ...), in any order. run_ue_py.ps1 -Script ...\set_build.py -ScriptArgs "all"places the sets. Rerun only when a layout changes.- Package as usual.
Options: one or more set names instead of all; --standins places flat boxes for meshes nobody has made (none are left); --front-offset 180; --dry-run; --dir-tag for a game without the settags patch.
Meshes
- Batches 2, 3, 4, 6, overnight 1 (
D:\studio-ai\renders\megajam\...) and the batch 15 keepers (D:\megajam\props\batch15\,game/props/batch15-manifest.json). - Made here (
meshgen/, intoD:\megajam\props\sets_made\):make_set_meshes.py(dirt and sawdust ground for the midway, a trampled field, a plank tent floor and its dirt yard, the beer tent shell: striped canvas roof on a ridge, back and gable walls, a scalloped valance over the open front),make_rope.py,raise_platform.py. Their textures are painted in numpy from the style-film-1 palette; the overnight 1 ground textures were not usable (FLUX drew little booths and lettering into them). tex_saturation.pycompares every base colour texture against the style-film-1 key art (batch 7 concepts): a texture whose 90th percentile saturation is over 1.25 x the key art's is toned down bymeshgen/desat_glb.py(only the base colour images are re-encoded) intoD:\megajam\props\sets_made\desat\, and meshes.json points at that copy. Nine were: beer mug, canvas awning, high striker, mirror maze facade, pennant pole, popcorn cart, teddy bears, sandbag, steel drum.- The show side:
FRONTin layout_tool.py for the meshes whose front is not local +Y (ticket booth, bar counter and mirror maze face local -Y; the armored car shows its broadside, local +X). Looked at from both sides on orthographic renders of the GLBs.
Checks
python layout_tool.py check: every item against the takes as they are (corridors, sight lines, props, reserved boxes, views).python tests/test_set_build.py: set_build.py against a stub of theunrealmodule.- Coverage, in the packaged game:
shots_run.ps1 -Pkg <build>plays each take with the perfect bot five times, with-KKShotAtat three moments and-KKShotModenormal (frame_stats.py numbers, and the pelvis log),mask_set(only KK_set actors over black, no post, sky or fog; the packaged build ignoresviewmode unlit, so they are lit, which a 2/255 threshold tolerates),mask_stand(the same without the ground covers),mask_all(dressing, take props and cast) andnocast(characters hidden: the set-only views, the only frames anyone looks at).python coverage.pygives each shot's share of dressing, standing dressing, props and cast, and bare floor or sky. Target: no shot over 40 % bare. Read the standing column too: the chase Steadicam looks down at the dirt, so the ground covers carry most of the chase's coverage.-KKCamAt=x,y,z,yaw,pitch[,fov];...puts shot i on camera i (the Stage 7 views).
Size budget (director, 2026-11-13: zip at or under 840 MB)
set_slim.py report|apply (meshes.json "budget" from layout_tool.py): base colour capped at hero 2048 (cart, gate, flatbed, armored car), dressing 1024, background 512; normal and occlusion/roughness maps at half that; Nanite keeps 100/35/20 % of the triangles; the mesh distance field at 1.0/0.35/0.2 of its default resolution. Every set mesh shares one of two Interchange parents (MI_Default_Opaque, MI_Default_Opaque_DS), so materials cost 0.2 MB in all.
Shared clip names (found here, fixed in director_build.py)
The director's baked clip copies were named <clip>__main<k> with no take in the name, so two takes that use the same clip at the same index overwrote each other's copy: in full5 the chase played the brawl's dust_off_s1 (the kid stood at the origin from 33.7 to 38.7 s) and the cannon replayed its climb at 20 s. They are now <clip>__<take>_main<k>. The old copies in Content/Motion/KimodoAligned are no longer referenced.
Set pass 2 (Fri 2026-11-13)
Batch 19 keepers (D:\megajam\props\batch19), strung lights and bunting, and a crowd, all through the same layout rules.
- Popcorn cart: the batch 19 cart c1 (240 x 135 x 219 cm, striped canopy) replaces batch 2's crumpled one under the same name, so the midway's spare cart and the chase take's cart follow. The vault cart is
cartat 0.43 all round (94 cm, the 93 the vault was timed over;chase.take.json). The old asset was deleted (set_build.py --reimport SM_popcorn_cart) and its toned GLB moved todesat\SM_popcorn_cart_batch2.glbso the layout tool does not pick it up for the new cart. - Sign boards (c1, blank gold face, 0.7 scale) stand on the roofs of the ring toss, cotton candy and duck gallery booths (
onitems, they follow their booth). - Carnival tent (c1, 396 x 402 x 500) at 1.5 x 1.5 x 1.0 = about 6 x 6 x 5 m replaces the three flat striped boxes on the cannon field and gets four neighbours, open fronts turned to the camera.
- String lights:
meshgen/make_set_spans.pymakes one 11.7 m festoon (wire at 4.0 m, 39 bulbs on ONE emissive slot, sockets on another).layout_tool.post_midwayhangs one between each pair of neighbouring light poles that were placed (they alternate sides, so the run zig-zags overhead), scaled along X to the exact pole to pole distance.set_build.pyputs the bulb slot onM_SetBulb(unlit, warm, 3.0 / 2.1 / 1.0). - Bunting: two 17.6 m lines gable wall to gable wall along the beer tent's ridge (4.9 m roof at 100 cm off the ridge, wire anchored at 4.6 m); 4 m and 6 m lines between the cannon field's flag poles (
post_cannonfield, pole top 147 cm to the -X side of the mesh centre, measured from the GLB). Strung items are centred meshes with their lowest point atZBASE(fromspans.json), so the layout sees them overhead (bottom over 300 cm) and they are exempt from the corridor and sight line rules like any overhead item. - Fewer crates on the midway: 7 down to 1 (the street slots and four booth-front clutter crates left out).
- Textures:
tex_saturation.pyflagged the cart, the sign and the three bunting meshes; toned withdesat_glb.py(0.55 / 0.5 / 0.55 / 0.6 / 0.55), all now under 1.25 x the key art. - GLBs with several parts import as one static mesh per NODE in Interchange, which cut the string lights down to their wire.
meshgen/ _glbmerge.pyrewrites the JSON chunk so all parts are primitives of one mesh. String lights, bunting and hats are not Nanite.
Crowd extras
Role crowd in the layouts, mesh name KK_CrowdExtra (no mesh: set_build.py spawns BP_Rex_final, the one shared MetaHuman body, face and skeleton; no new MetaHuman). 19 on the midway, 12 in the beer tent, 12 on the cannon field.
- Clips: the batch 19 loops idle_crowd_watch-s20, cheer_arms_up-s12, clap_steady-s13, eat_popcorn-s20, gasp_hands_to_face-s12, drink_from_cup-s4, retargeted onto Rex_final by
Scripts/crowd_retarget.ps1(same path as the longer takes; they land in/Game/Motion/Kimodo). They play in place from the clip itself: pelvis XY within 6 cm of the origin and heading 90 degrees (the body's own front) at frame 0, so actor yaw = facing - 90 like every other item. Mix per set inCROWD_MIX(drinkers in the tent, cheering on the field); each extra gets a random start offset inside its clip (start_s), a play rate of 0.93 to 1.07 and a height of 0.95 to 1.05. - Variants: four outfit colours (flat material instances on the costume's jacket, jeans and shirt,
CROWD_OUTFITS), five hats (none, Walt's flat cap, and boater, wide brim and derby frommeshgen/make_hats.py, authored in the head bone's frame like the cap and attached to theheadsocket), clean shaven or stubbled. A hat hides the hair groom under it. The faces are all Rex's: a crowd of cousins. - Faces: every extra has Rex's face, so the layout keeps each one at least 8 m (plan view) from every sampled camera position of its take (
CROWD_MIN_CAM, inplaceandcheck) and gives any extra within 16 m a brimmed hat (boater, wide brim or derby). - Placement: the same rules as every item (1 m corridors, sight lines, reserved boxes, take props, inside some camera's view); a spot that breaks one is nudged outward or dropped. The crowd stands out of the actors' corridors by construction, and
layout_tool.py checkcovers it. - Order of operations in
place_crowdmatters:set_editor_propertyon the actor or a groom reruns the Blueprint's construction script, which rebuilds Body and drops the clip and anything attached to it. Tags and groom visibility go first, the clip and the hat last. - Looking:
-KKShotMode=nocasthides every actor with a skeletal mesh and what is attached to one, so the crowd and its hats are gone from the set-only views. The other modes (normal, mask_set, mask_stand, mask_all) have the crowd in them: numbers only.
results.md
Source game/voice/results.md. Last git commit 2026-11-12. About 1,473 words.
Voice bake-off, tower 1
Takes with WER over 10% dropped. Reels loudness-matched to -20.0 LUFS; the LUFS column is the raw mean of the chosen takes. Aesthetics = mean of Audiobox PQ and CE (1 to 10). Cohesion = mean ECAPA similarity between the chosen takes (1.0 = same voice). Rank score is numbers only and cannot judge accent or character.
Walt
| rank | design | engine | lines ok | mean WER | LUFS | aesthetics | pitch Hz | cohesion | score | sounds like, from the numbers |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | onyx | Kokoro | 6/6 | 0.000 | -22.1 | 7.07 | 85.0 | 0.75 | 7.07 | deep (85 Hz), natural pace, holds one voice, aesthetics 7.1/10 |
| 2 | onyx | Chatterbox | 6/6 | 0.000 | -14.2 | 6.92 | 104.0 | 0.67 | 6.92 | baritone-range (104 Hz), fast, voice drifts a little, aesthetics 6.9/10 |
| 3 | mentor | Maya1 | 6/6 | 0.000 | -18.9 | 7.17 | 101.0 | 0.57 | 6.69 | baritone-range (101 Hz), slow, voice drifts a little, aesthetics 7.2/10 |
| 4 | fenrir | Kokoro | 6/6 | 0.000 | -18.7 | 6.88 | 141.0 | 0.74 | 6.4 | high for him (141 Hz), natural pace, holds one voice, aesthetics 6.9/10 |
| 5 | santa | Chatterbox | 6/6 | 0.000 | -15.6 | 7.16 | 196.0 | 0.74 | 5.66 | high for him (196 Hz), natural pace, holds one voice, aesthetics 7.2/10 |
| 6 | qwenref | Chatterbox | 6/6 | 0.000 | -19.8 | 7.01 | 225.0 | 0.77 | 5.51 | high for him (225 Hz), natural pace, holds one voice, aesthetics 7.0/10 |
| 7 | santa | Kokoro | 6/6 | 0.009 | -21.4 | 7.14 | 194.0 | 0.72 | 5.5 | high for him (194 Hz), natural pace, holds one voice, aesthetics 7.1/10 |
| 8 | brooklyn | Maya1 | 6/6 | 0.000 | -18.7 | 7.2 | 150.0 | 0.36 | 4.71 | high for him (150 Hz), natural pace, sounds like different people line to line, aesthetics 7.2/10 |
| 9 | gravel | Maya1 | 6/6 | 0.000 | -16.7 | 7.04 | 226.0 | 0.43 | 4.22 | high for him (226 Hz), slow, sounds like different people line to line, aesthetics 7.0/10 |
| 10 | mentor | Qwen3-TTS | 6/6 | 0.000 | -21.4 | 7.19 | 156.0 | 0.31 | 4.22 | high for him (156 Hz), slow, sounds like different people line to line, aesthetics 7.2/10 |
| 11 | mentor | VoxCPM2 | 6/6 | 0.000 | -22.7 | 7.04 | 149.0 | 0.3 | 4.22 | high for him (149 Hz), natural pace, sounds like different people line to line, aesthetics 7.0/10 |
| 12 | gravel | Qwen3-TTS | 6/6 | 0.000 | -19.8 | 6.95 | 254.0 | 0.4 | 3.95 | high for him (254 Hz), slow, sounds like different people line to line, aesthetics 7.0/10 |
| 13 | gravel | VoxCPM2 | 6/6 | 0.009 | -17.2 | 6.8 | 180.0 | 0.38 | 3.54 | high for him (180 Hz), natural pace, sounds like different people line to line, aesthetics 6.8/10 |
| 14 | brooklyn | Qwen3-TTS | 6/6 | 0.000 | -21.7 | 7.0 | 237.0 | 0.25 | 3.1 | high for him (237 Hz), slow, sounds like different people line to line, aesthetics 7.0/10 |
| 15 | brooklyn | VoxCPM2 | 6/6 | 0.000 | -19.3 | 6.99 | 246.0 | 0.22 | 2.91 | high for him (246 Hz), natural pace, sounds like different people line to line, aesthetics 7.0/10 |
| 16 | s3 | Dia | 5/6 | 0.012 | -14.5 | 6.17 | 132.0 | 0.21 | 2.47 | baritone-range (132 Hz), slow, sounds like different people line to line, aesthetics 6.2/10, only 5 of 6 lines passed the word check |
| 17 | s2 | Dia | 5/6 | 0.012 | -16.7 | 5.09 | 79.0 | 0.35 | 2.26 | deep (79 Hz), natural pace, sounds like different people line to line, aesthetics 5.1/10, only 5 of 6 lines passed the word check |
| 18 | s1 | Dia | 6/6 | 0.000 | -15.6 | 6.1 | 213.0 | 0.22 | 2.02 | high for him (213 Hz), natural pace, sounds like different people line to line, aesthetics 6.1/10 |
Rex
| rank | design | engine | lines ok | mean WER | LUFS | aesthetics | pitch Hz | cohesion | score | sounds like, from the numbers |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | qwenref | Chatterbox | 4/4 | 0.000 | -18.2 | 7.15 | 127.0 | 0.69 | 7.15 | baritone-range (127 Hz), natural pace, voice drifts a little, aesthetics 7.2/10 |
| 2 | adam | Kokoro | 4/4 | 0.000 | -17.9 | 6.87 | 126.0 | 0.73 | 6.87 | baritone-range (126 Hz), fast, holds one voice, aesthetics 6.9/10 |
| 3 | michael | Kokoro | 3/4 | 0.000 | -22.8 | 7.32 | 122.0 | 0.8 | 6.32 | baritone-range (122 Hz), natural pace, holds one voice, aesthetics 7.3/10, only 3 of 4 lines passed the word check |
| 4 | michael | Chatterbox | 3/4 | 0.000 | -15.5 | 7.06 | 132.0 | 0.66 | 6.0 | baritone-range (132 Hz), fast, voice drifts a little, aesthetics 7.1/10, only 3 of 4 lines passed the word check |
| 5 | liam | Kokoro | 3/4 | 0.000 | -21.3 | 7.0 | 116.0 | 0.81 | 6.0 | baritone-range (116 Hz), fast, holds one voice, aesthetics 7.0/10, only 3 of 4 lines passed the word check |
| 6 | adam | Chatterbox | 3/4 | 0.000 | -17.0 | 6.83 | 110.0 | 0.7 | 5.83 | baritone-range (110 Hz), fast, holds one voice, aesthetics 6.8/10, only 3 of 4 lines passed the word check |
| 7 | velvet | Maya1 | 4/4 | 0.000 | -18.9 | 6.97 | 105.0 | 0.46 | 5.83 | baritone-range (105 Hz), slow, sounds like different people line to line, aesthetics 7.0/10 |
| 8 | velvet | Qwen3-TTS | 4/4 | 0.000 | -21.2 | 7.17 | 122.0 | 0.38 | 5.55 | baritone-range (122 Hz), natural pace, sounds like different people line to line, aesthetics 7.2/10 |
| 9 | matinee | Maya1 | 4/4 | 0.000 | -18.1 | 6.98 | 135.0 | 0.39 | 5.27 | baritone-range (135 Hz), slow, sounds like different people line to line, aesthetics 7.0/10 |
| 10 | nervous | Maya1 | 4/4 | 0.000 | -18.7 | 6.64 | 97.0 | 0.39 | 5.08 | deep (97 Hz), slow, sounds like different people line to line, aesthetics 6.6/10 |
| 11 | nervous | Qwen3-TTS | 4/4 | 0.000 | -21.2 | 7.23 | 118.0 | 0.22 | 4.65 | baritone-range (118 Hz), natural pace, sounds like different people line to line, aesthetics 7.2/10 |
| 12 | matinee | Qwen3-TTS | 3/4 | 0.030 | -23.1 | 7.3 | 114.0 | 0.33 | 3.93 | baritone-range (114 Hz), natural pace, sounds like different people line to line, aesthetics 7.3/10, only 3 of 4 lines passed the word check |
| 13 | velvet | VoxCPM2 | 4/4 | 0.000 | -21.6 | 6.51 | 244.0 | 0.26 | 2.67 | light (244 Hz), natural pace, sounds like different people line to line, aesthetics 6.5/10 |
| 14 | matinee | VoxCPM2 | 4/4 | 0.000 | -17.3 | 6.64 | 294.0 | 0.18 | 2.32 | light (294 Hz), natural pace, sounds like different people line to line, aesthetics 6.6/10 |
| 15 | nervous | VoxCPM2 | 3/4 | 0.000 | -21.9 | 6.94 | 254.0 | 0.23 | 1.92 | light (254 Hz), natural pace, sounds like different people line to line, aesthetics 6.9/10, only 3 of 4 lines passed the word check |
| 16 | s2 | Dia | 4/4 | 0.000 | -15.3 | 5.37 | 197.0 | 0.15 | 0.87 | light (197 Hz), natural pace, sounds like different people line to line, aesthetics 5.4/10 |
| 17 | s1 | Dia | 2/4 | 0.042 | -19.5 | 4.17 | 137.0 | 0.55 | 0.73 | light (137 Hz), slow, voice drifts a little, aesthetics 4.2/10, only 2 of 4 lines passed the word check |
| 18 | s3 | Dia | 3/4 | 0.000 | -16.1 | 3.81 | 242.0 | 0.34 | -0.55 | light (242 Hz), slow, sounds like different people line to line, aesthetics 3.8/10, only 3 of 4 lines passed the word check |
fable-editbay.md
Source game/trials/fable-editbay.md. Last git commit 2026-11-12. About 422 words.
Trial: Fable 5.2 builds the edit bay (Thu Nov 12)
Don approved one controlled Fable 5.2 trial (chat, Thu 15:37). Launch after the 4 PM quota reset, never before.
Question
Does Fable 5.2 finish a large, self-contained game feature with fewer turns and fewer fix rounds than Opus 5.5, enough to justify about 2x the price per token?
Arms
- A (Fable 5.2 worker): the edit bay screen, brief below.
- B (Opus 5.5 worker): the call-sheet menu, a feature of similar size, the same brief shape and the same checks. It runs after A finishes, because the laptop is the only compiler.
Measures (recorded in game/devlog/trials.md)
- Wall time, tool calls and tokens from the agent result
- Compile attempts until the first clean build
- Checks passed on the first hand-in (build, cue tests, new unit tests, bot capture)
- Blind review: a fresh Opus reviewer scores both diffs against lab/STANDARDS.md without knowing which model wrote which
- Quota points used, from the usage meter before and after
Brief for arm A (edit bay)
Read first: game/DESIGN.md, game/STORY.md, Source/KidKatapult/Game/KKGame., Source/KidKatapult/CueGate/, game/takes/*.take.json, Scripts/build_editor.ps1, run_cue_tests.ps1, capture_bot.ps1.
Build the edit bay in the KidKatapult UE 5.9 C++ project:
- Record each attempt of a take: per-cue grade, timing offset, and which branch played (clean, flub, wreck).
- After a take, the edit bay screen lists the sections of that take. For each section the player picks which attempt goes in the film. Bloopers (flubs, wrecks) can go in on purpose. Rex's close-up is an extra angle per section, where the take has one.
- The choice is saved as a FinalCut (per scene, per section: attempt id, angle) that the premiere will play later. Keep FinalCut a plain data type with no UI dependency.
- Input: controller, touch and keyboard all work (studio rule). About 20 seconds of player time per scene, so keep it to a few choices.
- Text the player reads follows lab/WRITING.md (run studio-pipeline/tools/writing/tells.py on it). No em or en dashes.
Checks before hand-in:
- build_editor.ps1 is clean on the laptop
- run_cue_tests.ps1 passes, plus new automation tests for attempt recording and FinalCut selection
- A bot capture of the screen saved to D:\megajam\review\editbay\ (internal only, never shown to Don as greybox)
Rules: the laptop is the only compiler. Never run the editor beside a render. Never delete a loaded build folder; build into new folders. Commit by explicit paths in lab/megajam with a clear message. Stop and report if blocked; don't widen scope.