The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Wizard: data, weights and code
Raw video / condition pairs used to train the Wizard world model, the model weights, and the training and inference code. Every training window is 121 frames at 30 fps (about 4 s), cropped to 832×480.
The full training set has 50,115 windows (56 h). This repository holds the 26,444 of them that can be redistributed, plus 3,606 synthetic look-back windows that were added in the causal and distillation stages. The 23,654 WildWorld windows are not included (non-commercial licence, no redistribution), nor is anything derived from them.
| folder | source | windows | target video |
|---|---|---|---|
synthetic/ |
procedural renders: characters entering, co-existing, re-entering, fast turns | 9,386 | included (raw/) |
eldenring_combat/ |
Elden Ring boss fights (er 6,741, bk 1,141) |
7,882 | included (raw/) |
eldenring_exploration/ |
Elden Ring roaming (roam) |
3,404 | included (raw/) |
unreal/ |
fights simulated in Unreal Engine (ue) |
2,966 | included (raw/) |
synthetic_lookback/ |
procedural renders: the camera turns away from a character and back (lookback0 / lookback: first / second window of a 241-frame clip); causal and distillation stages only |
3,606 | included (raw/) |
scenefly/ |
SceneFly camera flights that return to earlier views (sfly 1,057, sfly_replay 1,344 reversed) |
2,401 | included (raw/), cut from upstream Orange-3DV-Team/SceneFly (Apache-2.0) |
worldedit/ |
world-edit clips generated with LTX-2.5 (wedit) |
405 | included (raw/) |
weights/ |
model weights, see weights/README.md |
||
code/ |
training and inference code, see code/README.md |
||
examples/ |
three ready-to-run inference cases with our outputs, see examples/README.md |
Reproducing the model
- Pre-encoded training data:
yuzhewu207/wizard-precoded— the exact training files (Wan2.2-VAE latents etc.) for all redistributable windows;python code/scripts/restore_precoded.py --root <dir>downloads and lays them out. - Base model:
Wan-AI/Wan2.2-TI2V-5B(WIZARD_WAN). - Code:
code/here (pip install -e code), one launch script per stage incode/scripts/; the weights each stage starts from are inweights/(seeweights/README.md). - Inference:
code/README.md→ Inference, with the ready-made cases inexamples/.
Data layout
Every folder has the same layout:
<folder>/
windows.jsonl one line per window: target video, caption, reference images, condition and geometry keys, origin
shards.json member file -> tar shard (look-up by key)
raw/shard-XXXXX.tar the target videos
cond_v7/main/shard-XXXXX.tar <key>.mp4 + <key>.npz the condition (proxy) video of the window
cond_v7/{unk2,trueunk,memunk}/ the same windows with parts of the scene marked unknown (only where such a variant was trained)
geo/shard-XXXXX.tar <key>.npz per-window cameras and depth
refs/ character reference images (fights)
memory/ 1-fps memory-bank metadata and retrieval indices (exploration, SceneFly, world edit)
All bulk files are WebDataset-style tar shards (about 2 GB each). The formats of the condition video, its .npz and the geometry file are the same in every folder and are described below under the synthetic part.
Target videos of the real sources
windows.jsonl gives "video": "<folder>/raw/<key>.mp4", a 121-frame clip, and "origin", where the window came from (recording, start frame).
eldenring_combat/,unreal/: the pre-cut 121-frame, 30 fps windows the training latents were encoded from (1024×576, or 960×540 forbk), copied unchanged.eldenring_exploration/: frames[start_frame, start_frame + 121)of the recording, re-encoded (x264, CRF 14, 960×540, 30 fps).worldedit/: the generated clip as it is (1024×576, 121 frames, stored at 24 fps with the generated audio track; training used the 121 frames).originholds the LTX prompt, seed and edit type. Captions are the single roaming caption used in training.scenefly/: frames[start_frame, start_frame + 121)of the upstream video given inorigin, scaled from 1920×1080 to 960×540 and re-encoded (x264, CRF 14). Windows with"reversed": trueare stored played backwards, as they were trained. Seescenefly/LICENSE_SceneFly.md.
Training scales every target video to height 480 and crops 832 columns starting at x = 10.
Memory
memory/*.tar holds, per recording segment, the 1-fps memory-bank frames (f frame number, t, depth, pos, R, K); the picture of a memory frame is frame f of the source video. The *_mem_index*.json files give, for every window and every chunk of 4 latent frames, the retrieved memory frames as [[frame, coverage gain], ...]. memroam_rep_keys_both30.json lists the roaming windows where at least 30% of the cells can only be seen in memory (repeated 8× in training).
Synthetic renders (synthetic/, synthetic_lookback/)
9,386 windows rendered procedurally: characters entering the frame, several characters sharing a scene, leaving and re-entering, and fast turns under a follow camera.
| group | windows | content | raw clips |
|---|---|---|---|
e1o / e1n / e1x |
3,060 / 1,910 / 992 | one character enters (straight / oblique / orbit, leaves and re-enters) | raw/entry1_v2, raw/entry1_v3, raw/entry1_v3x |
e2o / e2n |
1,792 / 1,424 | two characters enter, each bound to its own reference image | raw/entry2_v2, raw/entry2_v3 |
turn |
208 | running and turning back | raw/turn_r2 |
synthetic/
windows.jsonl one line per window: raw clip, caption, reference images, condition and geometry keys
shards.json member file -> tar shard (look-up by key)
raw/<set>/shard-XXXXX.tar <clip>/rgb.mp4, first_frame.png, ref*.png, meta.json, shot.npz, slots.npz, report.json
cond_v7/main/shard-XXXXX.tar <key>.mp4 + <key>.npz the condition (proxy) video of the window
cond_v7/{unk2,trueunk,memunk}/ alternative conditions with parts of the scene marked unknown (only some windows)
geo/shard-XXXXX.tar <key>.npz per-window cameras and depth
credits/ attribution of the 3D assets
All bulk files are WebDataset-style tar shards (about 2 GB each).
synthetic_lookback/ has the same layout; its clips are named <record>_<0|1> (the two 121-frame windows of one 241-frame clip, the second starting at the first one's last frame), windows.jsonl adds record, order and memory_prev_windows (the window whose frames serve as memory), and the scenes are procedural terrain with Poly Haven props instead of flat ground.
How a pair is formed
A line of windows.jsonl:
{"key": "69eb3dab907d5fc7", "group": "e2n",
"source": {"dir": "synthetic/raw/entry2_v3/t232_000030", "video": "rgb.mp4", "start_frame": 0, "n_frames": 121},
"caption": "...",
"refs": ["synthetic/raw/entry2_v3/t232_000030/ref_A.png", "synthetic/raw/entry2_v3/t232_000030/ref_B.png"],
"cond_v7": {"main": "synthetic/cond_v7/main/69eb3dab907d5fc7"},
"geo": "synthetic/geo/69eb3dab907d5fc7.npz"}
- Target video:
<source.dir>/rgb.mp4, the whole clip is the 121-frame window. Rendered at 1024×576; training scales it to height 480 and crops 832 columns starting at x = 10. - Condition:
<cond_v7.main>.mp4(832×480, 121 frames) with<cond_v7.main>.npz. - Reference images:
refs(480×480 in training). - Caption: the caption used in training, unchanged (Chinese).
Use shards.json to find the tar that holds a given member.
Condition video
<key>.mp4 is H.264 yuv444p10le, CRF 12, full-range BT.709 (decode as full range). It is the Wan2.2-VAE decode of the exact condition latent the model was trained on. Channels encode geometry, not appearance:
- R: world height. B: depth along the camera axis,
1 - clamp((z - 0.1) / 79.9). - G: for the scene, how reliable the geometry is: sky 0, unknown 38 (R = B = 0, dark green), coarse 77, exact 153–255 (modulated by Lambert shading under a fixed world light). For characters, G separates the player from the others and encodes facing.
- Character skeletons are drawn on top with a 2 px black outline.
<key>.npz holds the lossless per-cell arrays on the 104×60 token grid: dep [121, 60, 104] float16 (metres, 0 on characters and sky), tier [121, 60, 104] uint8 (0 sky, 1 unknown, 2 coarse, 3 exact), and meta (JSON string: source and variant).
Geometry
geo/<key>.npz: depth [121, 60, 104] float16 scene depth in metres (completed behind the characters); pos [121, 3] and R [121, 3, 3] camera (rows of R are right / up / forward, y up, metres); K [4] = fx, fy, cx, cy on the 104×60 grid of the 832×480 crop; origin [3] the player's root at the first frame; hp the player's height; actor [121, 60, 104] cells covered by characters.
Raw clip files
meta.json records the assets, motions and timing of each character (which frames it is in view, entry frame, visible area). shot.npz holds per-frame skeletons and cameras, slots.npz the image region of each character. Absolute paths inside meta.json refer to the machine the data was rendered on.
Licence and attribution
Research use only.
- Elden Ring recordings: gameplay footage; the game and its content are © FromSoftware / Bandai Namco. Released for non-commercial research only.
- Unreal Engine fights: recorded in scenes we assembled ourselves.
- SceneFly: Apache-2.0, Orange-3DV-Team.
scenefly/raw/holds windows cut from their videos (changes listed inscenefly/LICENSE_SceneFly.md); the depth, cameras, conditions and window indices were computed by us. - World-edit clips: generated with LTX-2.5 from first frames of the recordings above; subject to the LTX model licence.
- Synthetic characters: 140 CC-BY models from Sketchfab; author, licence and link of each are listed in
credits/sketchfab_credits.json. Further characters come from KayKit and Quaternius (CC0). Some models ship with Mixamo animations, which are subject to Adobe's Mixamo terms. A few models are fan-made renditions of third-party characters; the rights to those characters remain with their owners. - Motions: part of the synthetic clips are driven by motions retargeted from game animation data; the rights to the original animations remain with their owners.
- Environments: Poly Haven HDRIs and models (CC0); eight stylised panoramas are AI-generated.
credits/synth_assets_used.jsonlists every asset path used in the synthetic renders and how often.
If you are a rights holder and want something removed, please open a discussion on this repository.
- Downloads last month
- 87