wizard-precoded / README.md
yuzhewu207's picture
README: point to ww_index
69282c2 verified
|
Raw History Blame Contribute Delete
5.92 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 yuzhewu207/wizard code/ (package 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, rewrites paths (~0.85 TB free disk)
export WIZARD_DATA=/path/to/wizard_data
export WIZARD_WAN=/path/to/Wan2.2-TI2V-5B                        # huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ...
cd code   # from yuzhewu207/wizard
TEACHER=../weights/causal_teacher/dit_step11000.safetensors OUT=runs/consistency bash scripts/distill_consistency.sh

restore_precoded.py (in this repository) downloads precoded/ shard by shard (resumable, each tar is deleted after extraction), extracts everything into one root, copies precoded_meta/ over it, and rewrites the absolute paths the data was built with (/data/yuzhewu/datasets) in symlinks and row files to your root. --verify checks every file against the sha256 manifests; --dirs wan_v7,wan_mix,... restores a subset.

What is where

precoded/<folder>/shard-NNNNN.tar      members are paths relative to the data root, e.g. wan_v7/cond7/clip_<key>.pt
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
wan_v7 --data $WIZARD_DATA/wan_v7 v7_rows.jsonl (the row file), cond7/ proxy-video latents (+ cond7_unk2/, cond7_trueunk/, cond7_memunk/ unknown-region variants), gbuf2 → wan_v6/gbuf2
wan_v6_gbuf2 wan_v6/gbuf2/ per-window depth (60×104), cameras, character masks
wan_mix, wan_mix5 wan_mix5/clips/clip_<key>.pt (symlinks into wan_mix/clips and the folders below), wan_mix5/text_table.pt, neg_emb.pt target latents, reference latents + slots, character cover per token, prompt
wan_entry_v3, wan_entry_v3x, wan_turn_g, wan_roam_unarmed, wan_lookback, wan_lookback_geo through wan_mix5/clips and --extra_rows synthetic entry / re-entry / turn / look-back windows
wan_memroam wan_memroam/ Elden Ring roaming windows, their memory banks (bank/<segment>/), mem_index*.pt, rep_keys_both30.json
wan_sfly, wan_sfly_replay, wan_sfly_replay_test --extra_rows, wan_sfly/ banks SceneFly camera flights (forward and reversed) + memory banks
wan_wedit 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 four stages read this layout, with the launch scripts of code/scripts/ (8 GPUs, FSDP, one window per GPU):

stage script start from released result
0 bidirectional train_bidirectional.sh Wan2.2-TI2V-5B weights/bidirectional/dit_step35000
1 causal teacher train_causal.sh stage 0 weights/causal_teacher/dit_step11000
2 consistency distillation distill_consistency.sh stage 1 —
3 DMD distill_dmd.sh stage 2 EMA (student), stage 0 (real score) weights/causal_dmd/ema_step480 (earlier chain)

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 release 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).