HyebinKim/re10k-ztoz-tokens

RE10K ZipSplat Token Pairs (Z, Z′)

Cached latent-token pairs produced by the released ZipSplat encoder on the RealEstate10K training split. Each record stores the scene tokens before test-time optimization (TTO) and the same tokens after 50 TTO steps, so a model can be trained to map Z → Z′ directly in token space instead of running TTO at inference time.

The tokens are the only thing stored here — no images, no camera poses, no depth. Everything needed to render is either in the token tensors or reproducible from RealEstate10K itself.

Contents

Path	Description
index.json	{scene_key: [{group, shard, context_indices, target_indices, seed, num_views}]}, one entry per draw
shards/shard-0000.h5 … shard-0063.h5	64 HDF5 shards, 713.7 GB total, one HDF5 group per draw

• 60,962 scenes, one view draw per scene (group name <scene_key>__s0).
• Shard assignment is sha256(scene_key) % 64, so a scene's shard is reproducible without reading index.json.
• Datasets are gzip level 1 compressed.

Record schema

Every HDF5 group holds three datasets, all float16:

Dataset	Shape	Meaning
scene_tokens_before	[1944, 1536]	Z, the frozen encoder output
scene_tokens_after	[1944, 1536]	Z′, after 50 TTO steps
color_feats	[1944, 128]	color-skip features, never optimized, so before and after are identical and it is stored once

1944 = 6 context views × 324 tokens per view. Token blocks are concatenated per view, and context view 0 defines the normalization frame, so the order in context_indices is meaningful.

Group attributes: context_indices, target_indices, sup_indices, num_views, draw, seed, tto_steps, tto_lr, eval_use_priors, context_mode, supervision, complete.

How the pairs were produced

1. View sampling. ZipSplat's own training-time bounded view sampler (split="train") with the RealEstate10K settings from splatfactory/configs/data/composed_252.yaml: 6 context views, 4 target views, min_distance_between_context_views=35, gap_multiplier=45, max_distance_between_consecutive_views=90. Context order is shuffled the way the training dataloader shuffles it, so the reference view is a random one of the six. Draws are seeded from sha256(scene_key) + draw index, so they reproduce exactly.
2. Filtering. The same two filters the training dataloader applies: at least 36 frames per scene, and a max/median consecutive camera step ratio of at most 10 over the frame range the draw spans. 5,071 of the 66,033 source scenes were dropped by these filters.
3. Normalization. Poses relative to context view 0, translations and depths divided by the scene scale, near = 0.1 / scale, far = 100 / scale.
4. Encoding. Frozen released ZipSplat checkpoint in .eval() mode with eval_use_priors=True (depth and pose priors injected), image resize 252, edge_divisible_by=14.
5. TTO. AdamW on the scene tokens only (optimize_color=False), L1 photometric loss with no LPIPS term, 50 steps at lr=3e-3, betas=(0.9, 0.95), weight_decay=0.05, gradient-norm clip 1.0. The photometric loss is supervised on the draw's own 6 context views.

Usage

import json
import h5py
from huggingface_hub import hf_hub_download

index = json.load(open(hf_hub_download("HyebinKim/re10k-ztoz-tokens", "index.json", repo_type="dataset")))
scene_key, entries = next(iter(index.items()))
entry = entries[0]

path = hf_hub_download("HyebinKim/re10k-ztoz-tokens", entry["shard"], repo_type="dataset")
with h5py.File(path, "r") as f:
    group = f[entry["group"]]
    before = group["scene_tokens_before"][:]   # (1944, 1536) float16
    after = group["scene_tokens_after"][:]     # (1944, 1536) float16
    color = group["color_feats"][:]            # (1944, 128)  float16

Each shard is about 11 GB, so download only the shards you need rather than the whole repository. To stream the full set, hf download HyebinKim/re10k-ztoz-tokens --repo-type dataset --local-dir <dir>.

Reproducing

The export is deterministic given the same RealEstate10K tar shards and the released ZipSplat checkpoint: seeds come from the scene key, and eval_use_priors, the TTO schedule and the view sampler are all pinned. Regenerating the full set takes roughly 3.5 hours on 4 GPUs at about 0.8 s per draw per worker.

Licensing and provenance

These tokens are derived from RealEstate10K, which is distributed by Google under its own terms for non-commercial research use. Use this data under those terms, and cite the original dataset:

@article{zhou2018stereo,
  title   = {Stereo Magnification: Learning View Synthesis using Multiplane Images},
  author  = {Zhou, Tinghui and Tucker, Richard and Flynn, John and Fyffe, Graham and Snavely, Noah},
  journal = {SIGGRAPH},
  year    = {2018}
}

If you use the tokens themselves, please also cite ZipSplat, whose encoder and TTO procedure produced them.