teawhite/GameCrafter-Data-v8

GameCrafter v8 Paired Data

Version 8 of the paired training data for an action-conditioned 2.5D game world model whose control signal is a per-pixel numeric semantic map. It follows GameCrafter-Data-v4 and keeps its camera, frame format, action alignment and index files. The worlds and fights are new:

• each run's world mixes objects and materials from several of the 10 scenes;
• the RGB pass has a higher-quality art setup;
• player and boss are new action-RPG characters with weapons, combo movesets and signature moves.

Each frame stores:

• the rendered RGB frame;
• a pixel-aligned semantic id map (one integer per pixel);
• a structural "whitebox" render (scene geometry plus articulated capsules for the characters);
• the player's 5-dim action (W, A, S, D, attack);
• camera, actor, character, appearance and event metadata.

What changed since v4

Versions 5–7 were internal steps; v8 is the first public release since v4 and includes all of them.

1. Layout remix (v5).
   • Each run rearranges the host scene: walls, archways, huts, groves, hedgerows and rock formations are added or relocated, and props are shuffled.
   • Some recipes mirror the world in X and / or Y.
   • The remixed objects are also written into the whitebox.
2. Cross-scene worlds (v6).
   • Every run transplants buildings, structures, props, rocks and plants from 2–3 other scenes ("donors") into the host scene.
   • Some recipes also swap the host's materials for those of another scene.
   • Five recipes (C1–C5, below) are rotated.
3. Art pass on the RGB (v7).
   • Lumen AO, bloom, a light vignette and an exposure offset.
   • Real rock and prop meshes replace the primitive procedural ones (pools of 53 rocks and 33 props).
   • Small clutter plants and stones are scattered on the ground.
4. New characters and fights (v8).
   • Characters. 8 player characters and 9 bosses, generated as textured 3D models and auto-rigged on the UE5 mannequin skeleton. Each carries one of 10 hand-held weapons, such as a katana or a war hammer.
   • Movesets. Locomotion, a 3-hit light combo, a heavy attack, dodge, guard, stagger, hit reaction and death, all retargeted from CC0 and UE clips. Bodies lean procedurally and look at the opponent.
   • Signature moves. Whirlwind (two spinning turns that re-hit every 0.28 s) and leap slam (a jump arc onto the opponent with an area slam and a ground shockwave).
   • Hit feedback. Hit stop, knockback scaled by relative size, an impact burst, and a short rim-only hit flash on the struck character. The FX meshes are RGB-only and never appear in the labels or the whitebox.
   • Readability. Characters are rendered larger than in v4, with an RGB-only outline, rim light and a saturation / contrast boost. The background is slightly toned down so the fighters stand out. Labels are unchanged.
   • Image quality. The RGB is rendered at 2×2 resolution and box-filtered in linear light; labels, depth and whitebox stay native and pixel-aligned. Character and weapon textures are forced fully resident at full resolution.
   • Theme-aware set dressing. Donor objects and props only come from scenes with a compatible look, are placed next to walls and structures, and come in small clusters.

At a glance

Scenes	10 UE 5.8 scenes, isometric orthographic camera (2.5D)
Runs	75 runs of 5 min (4,800 frames each)
Frames	360,000 at 16 fps (6.25 h of game time)
Resolution	1280 × 704
Characters	8 player and 9 boss models, 10 weapons
Fights	15,466 fight blocks; 742 whirlwinds and 739 leap slams
Dynamic events	3,525 (spawn 1,250 / replace 1,145 / delete 1,130)
Ready-made 81-frame clips	4,158, of which 3,069 (74%) contain at least one event
Appearances	6 (default, golden, overcast, snow, autumn, palette)
Label check	0 invalid semantic pixels across the whole set
Action check	0 key changes inside any 0.25 s block
Size	118 GB

Previews: previews/v8_showcase_part1.mp4 and previews/v8_showcase_part2.mp4 show the busiest ~11 s fight window of each scene's first run at native resolution.

Runs and recipes

Runs are named P8C<recipe>_<appearance>_s<bot seed>, for example P8C2_golden_s5802.

• The bot seed is scene × 1000 + 800 + round.
• Rounds rotate through the recipes: round 1 and 6 → C1, 2 and 7 → C2, 3 and 8 → C3, 4 → C4, 5 → C5.
• Odd rounds use the default appearance. Even rounds rotate through golden, overcast, snow, autumn and palette.
• Scenes 01–05 have 8 runs and scenes 06–10 have 7.

Recipe	Runs	World
C1	20	objects from two donors + light remix
C2	20	two donors + full material swap from a third scene + remix, X mirror
C3	15	two donors (strength 1.2) + 60% material swap from one of them, Y mirror
C4	10	two donors + full material swap from a third scene + remix, XY mirror
C5	10	heavy transplant (strength 1.6) from three donors + light remix

Per-run details are in index/capture_result.json: donors, transplanted object counts, mirror flags, remix counts and appearance parameters. summary_v8.csv lists every run.

Appearances

Appearance	Runs	Hours
default	40	3.33 h
golden	7	0.58 h
overcast	7	0.58 h
snow	7	0.58 h
autumn	7	0.58 h
palette	7	0.58 h

The appearances work as in v4. There is no night in this release.

Per-frame contents

Each frame is one WebDataset sample inside a tar shard. The sample key is the 6-digit frame index within its run.

File in tar	Content
{frame}.jpg	RGB, 1280×704, JPEG q85 (rendered at 2560×1408 and downsampled)
{frame}.ids.png	Semantic id map, single-channel uint8, lossless, pixel-aligned with the RGB
{frame}.whitebox.jpg	Whitebox render (structural scene geometry plus articulated capsules for the characters), JPEG q95
{frame}.json	Per-frame metadata (same fields as v4)

The per-frame json has the same fields as v4, including:

• world_to_clip and the camera pose;
• actors, with role, class id, state, health, position and velocity;
• action, with keys, block, bot_mode and player_stamina;
• characters, with model, material variant, weapon, scale, height and capsule;
• events, appearance and label_invalid_pixels.

Character model names are now the v8 characters, for example v8_samurai or v8_golem. material_variant also records the weapon, for example mat2|w_katana.

Semantic classes

id	class	id	class	id	class
0	sky / background	3	architecture	6	prop
1	terrain	4	rock	10	player
2	paving (roads, floors)	5	vegetation	11	boss

Pixel share, measured on 600 frames spread over all runs:

Class	Share
paving	31.9%
architecture	23.6%
vegetation	20.2%
terrain	18.6%
prop	2.5%
rock	2.4%
boss	0.58%
player	0.37%

The characters cover 2–3 times as many pixels as in v4.

Actions and clips

Same as v3 and v4:

• actions.npy has shape (N, 5) with columns [w, a, s, d, attack]. Row i holds the keys held from frame i to frame i+1.
• Decisions are only made on frames with i % 4 == 0, so keys never change inside a 0.25 s block. actions_block_0.25s.npy has shape (N/4, 5), one action prompt per block.
• clips_81.json lists 81-frame clips (start % 4 == 0, stride 80) that never cross an episode or camera change. Each clip lists its events with frame offsets.

Characters

Player model	Share of frames	Boss model	Share of frames
v8_merc	14.1%	v8_golem	19.2%
v8_rogue	14.0%	v8_demon	12.3%
v8_samurai	13.2%	v8_oni	11.1%
v8_ranger	12.8%	v8_skeleton2	10.4%
v8_viking	12.3%	v8_automaton	10.0%
v8_nomad	12.1%	v8_troll	9.7%
v8_paladin	11.1%	v8_minotaur	9.4%
v8_knight	10.3%	v8_orc	9.3%
		v8_deathknight2	8.7%

• Every episode picks a new player and a new boss, a material variant, a weapon (about 6% of armed characters fight unarmed for the episode), a skin recolour and a scale.
• Heights, p5–p95: player 225–259 cm, boss 298–386 cm. The capsule follows the model and scale, and movement speeds scale with size.
• Player states over the whole set include move, idle, hit, death, light1–light3, whirlwind and leap_slam. The per-frame state is in actions.csv and frames.jsonl.

Dynamic events

The events work as in v4: objects are deleted, replaced or spawned while an episode is running, and the RGB frame and the semantic map change on the same frame. All changed objects are restored when a new episode starts.

Action	Class	Count
removed	vegetation	1,534
added	rock	1,109
added	prop	867
added	vegetation	419
removed	prop	384
removed	rock	273
removed	architecture	84

Behaviour

Share of frames by bot mode over the whole set:

Bot mode	Share
explore	38.7%
fight	17.2%
seek_boss	12.8%
unstick	6.0%
pause	5.4%
dead	5.3%
detour	3.9%
retreat	3.7%
backtrack	3.1%
wallhug	1.8%
disengage	1.1%
push_obstacle	0.9%

Frames with no key held: 19.5%.

Scene	Runs	Episodes	Clips	Events	Fight blocks	Whirlwinds + leap slams	Fight share*
01 Sunlit Mediterranean Courtyard	8	61	447	398	1,068	92	13.9%
02 Autumn Woodland Shrine	8	58	446	396	1,360	116	17.5%
03 Desert Caravanserai	8	105	428	336	3,838	362	48.1%
04 Snowbound Mountain Outpost	8	60	448	397	1,460	124	18.4%
05 Japanese Spring Sanctuary	8	57	451	399	1,356	126	17.0%
06 Overgrown Tropical Water Temple	7	61	389	336	1,248	146	18.2%
07 Steampunk Foundry District	7	49	396	352	1,010	68	14.2%
08 Gothic Cathedral Quarter	7	51	393	269	1,000	87	13.2%
09 Dense Cyberpunk Neon District	7	86	376	320	1,567	213	23.3%
10 Underwater Research Habitat	7	68	384	322	1,559	147	23.8%

* Share of frames in which the bot is fighting or retreating.

Repository layout

data/<scene>/<run>/frames-000000.tar     frames 0–1599 of the run (WebDataset shard)
data/<scene>/<run>/frames-001600.tar     frames 1600–3199
data/<scene>/<run>/frames-003200.tar     frames 3200–4799
data/<scene>/<run>/index/
    actions.npy                (N,5) uint8 [w,a,s,d,attack]
    actions_block_0.25s.npy    (N/4,5) one action prompt per 0.25 s
    episodes.npy               (N,) episode id
    actions.csv                per-frame actions, bot mode, player/boss state and HP, character model and scale, event count
    frames.jsonl               compact per-frame metadata
    clips_81.json              81-frame training clips with their events
    events.json                all dynamic events of the run
    meta.json                  run statistics (appearance, cross-scene recipe, events, behaviour, characters)
    capture_result.json        capture settings, appearance, cross-scene / remix details, v8 character and fight counts
    scene_actors.json          static-mesh inventory of the host scene before augmentation
summary_v8.csv / summary_v8.json   statistics for all 75 runs and totals
previews/                          showcase videos
thumbs/                            one thumbnail per run (RGB | semantic | whitebox)

Loading a clip

import io, json, tarfile
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download

REPO = "teawhite/GameCrafter-Data-v8"
scene, run = "05_Japanese_Spring_Sanctuary", "P8C1_default_s5801"
get = lambda p: hf_hub_download(REPO, f"data/{scene}/{run}/{p}", repo_type="dataset")

clips = json.load(open(get("index/clips_81.json")))
blocks = np.load(get("index/actions_block_0.25s.npy"))
c = clips[10]; s, e = c["start"], c["end"]          # 81 frames; c["events"] lists events inside the clip

def frames(lo, hi):                                   # read frames [lo, hi) from the 1,600-frame shards
    out = {}
    for shard in range(lo // 1600 * 1600, hi, 1600):
        with tarfile.open(get(f"frames-{shard:06d}.tar")) as tf:
            for m in tf:
                key, ext = m.name.split(".", 1)
                if lo <= int(key) < hi:
                    out.setdefault(int(key), {})[ext] = tf.extractfile(m).read()
    return [out[i] for i in range(lo, hi)]

clip = frames(s, e)
rgb = np.stack([np.asarray(Image.open(io.BytesIO(f["jpg"]))) for f in clip])         # (81,704,1280,3)
ids = np.stack([np.asarray(Image.open(io.BytesIO(f["ids.png"]))) for f in clip])     # (81,704,1280) uint8
meta = [json.loads(f["json"]) for f in clip]                                          # characters, events, ...
act = blocks[s // 4 : s // 4 + 20]                                                   # (20,5) one prompt per latent group

Capture details

• Unreal Engine 5.8, isometric orthographic camera following the player, with a random camera yaw, pitch and ortho width per episode.
• Offline fixed-timestep simulation at 48 Hz, keeping every 3rd step: 16 fps of game time independent of wall-clock speed. The 6.25 h of data took about 19 h of capture time (about 15 minutes per run), next to another Unreal capture on the same GPU.
• Each run waits for shader and asset compilation to go idle and warms up before the first saved frame. Every frame's labels are checked for invalid pixels.

Known limitations

• Same 10 host scenes. All new variety comes from transplanting and remixing objects and materials between them.
• No night. This release has no night appearance; night was not tested with the v8 look.
• Scene 03 is fight-heavy (about 48% of frames), because its boss spawns next to the main route.
• RGB-only stylisation. The outline, rim light, hit flash and impact FX exist only in the RGB. The semantic map and the whitebox show the plain characters.
• Uneven run counts. Scenes 06–10 have one run fewer than 01–05, because the capture stopped at a fixed time.