wizard-precoded / README.md
yuzhewu207's picture
README: stage scripts as trained
c30fb96 verified
|
Raw History Blame Contribute Delete
5.97 kB
metadata
license: other
license_name: research-only
pretty_name: Wizard pre-encoded training data
tags:
  - world-model
  - video-generation
  - game
size_categories:
  - 10K<n<100K

Wizard: pre-encoded training data

This is the exact pre-encoded data the Wizard world model was trained and distilled on: Wan2.2-VAE latents of the target and proxy videos, UMT5 text embeddings, per-cell depth/cameras, character cover maps, reference-image latents, 1-fps memory banks and retrieval indices. Use it with the training code in yuzhewu207/wizard (code/) to run every training stage directly, without re-encoding anything.

windows 30,050 rows (65,124 samples per epoch with the released sampling weights) in 15 groups
size 804 GiB in 89 tar shards (~10 GiB each) + 0.3 GiB of row files and text tables
code wyzardWu/wizard27-clean (package wizard; mirrored in code/ of yuzhewu207/wizard)
raw form of the same windows yuzhewu207/wizard (videos, proxy videos, geometry)
not included the WildWorld windows (group ww, 23,654 rows of the full 47,292): non-commercial licence, no redistribution. ww_index/ has the window list, captions, lags and scripts to rebuild them from AlayaLab/WildWorld

Quick start

pip install huggingface_hub
python restore_precoded.py --root /path/to/wizard_data          # downloads, extracts, renames folders, rewrites paths (~0.85 TB free disk)
export WIZARD_DATA=/path/to/wizard_data
export WAN_CKPT=/path/to/Wan2.2-TI2V-5B                          # huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ...
git clone https://github.com/wyzardWu/wizard27-clean && cd wizard27-clean && pip install -e .
huggingface-cli download yuzhewu207/wizard --repo-type dataset --include "weights/causal_teacher/*" --local-dir hf_wizard
bash scripts/train/3_consistency.sh hf_wizard/weights/causal_teacher/dit_step11000.safetensors runs/consistency

What is where

precoded/<folder>/shard-NNNNN.tar      members are paths relative to the data root (build-time folder names; the restore script maps them)
precoded/<folder>/shard-NNNNN.tar.tsv  manifest: size  sha256  path   (symlinks: "symlink-><target>")
precoded_meta/                         row files and text tables, WildWorld rows and WildWorld-only captions removed
restore_precoded.py                    download + extract + path fix + verify
ww_index/                              how to rebuild the WildWorld windows from upstream (see ww_index/README.md)
folder read by the code as content
windows --data $WIZARD_DATA/windows rows.jsonl (the row file), proxy/ proxy-video latents (+ proxy_unk2/, proxy_trueunk/, proxy_memunk/ unknown-region variants), shield/, geometry → ../geometry
geometry windows/geometry/ per-window depth (60×104), cameras, character masks
latents, samples samples/clips/clip_<key>.pt (links into latents/clips and the folders below), samples/text_table.pt, neg_emb.pt target latents, reference latents + slots, character cover per token, prompt
synth_entry, synth_entry_orbit, synth_turn, roam_unarmed, synth_lookback, synth_lookback_geo through samples/clips and --extra_rows synthetic entry / re-entry / turn / look-back windows
roam_memory roam_memory/ Elden Ring roaming windows, their memory banks (bank/<segment>/), mem_index*.pt, rep_keys_both30.json
scenefly, scenefly_replay, scenefly_replay_test --extra_rows, scenefly/ banks SceneFly camera flights (forward and reversed) + memory banks
worldedit WEDIT=1, --extra_rows LTX-2.5 world-edit windows + banks

Groups (rows): er 6,741 · roam 3,404 · e1o 3,060 · ue 2,966 · e1n 1,910 · lookback 1,803 · lookback0 1,803 · e2o 1,792 · e2n 1,424 · sfly_replay 1,344 · bk 1,141 · sfly 1,057 · e1x 992 · wedit 405 · turn 208.

Reproducing the training stages

All stages read this layout, with the launch scripts in scripts/train/ of the code repository (8 GPUs, FSDP, one window per GPU); the weights are in weights/ of yuzhewu207/wizard:

stage script start from released result
1 bidirectional teacher 1_bidirectional.sh Wan2.2-TI2V-5B bidirectional/dit_step35000
2 causal teacher 2_causal.sh stage 1 causal_teacher/dit_step11000
3 consistency distillation 3_consistency.sh stage 2 (being re-run on stage 2; an earlier run is causal_cd/ema_step10000)
4 DMD 4_dmd.sh stage 3 EMA (student), stage 1 (real score) (being re-run; an earlier run is causal_dmd/ema_step480)

Checked before release on the full data (with the WildWorld windows): stage 2 started from causal_teacher/dit_step11000 reproduces the original training log value by value (rows 53704 -> 88778; step 10 cd 0.00018 sig 0.694->0.673 gnorm 0.02; step 20 cd 0.00068 vs 0.00069), at 16–20 s/step on 8×H200 with 38 GB per GPU. On this public part, the released data loaders read every group (375 random samples across all 15 groups, no missing file).

Without the WildWorld windows the training distribution differs from the paper's (they were about half of the windows), so expect close but not identical results; ww_index/README.md describes how to rebuild them from the upstream dataset and add them.

Licence

Research use only. Elden Ring material © FromSoftware / Bandai Namco; SceneFly Apache-2.0 (Orange-3DV-Team); world-edit clips generated with LTX-2.5 (LTX licence); synthetic characters CC-BY / CC0 (credits in yuzhewu207/wizard).