SceneForge: Structured World Supervision from 3D Interventions

Dataset release for the SceneForge paper (arXiv:2605.14399). Generated from Infinigen indoor scenes and Blender (Cycles). Exact counts are in stats.json.

Contents

Folder	What it is	Size
SceneForge-S/	Raw single-view corpus, packed as 10 tar shards (S-0000.tar …); each room has full_scene/ + object-linked layers/	2,407 rooms, 25,568 object layers
SceneForge-M/	Raw multi-view corpus, packed as 1 tar shard (M-0000.tar); one folder per room, one sub-folder per camera setting	517 rooms, 2,771 scene instances, 34,863 image layers
sceneforge-removal/sceneforge-removal-s/	SceneForge-Removal (S): orig/mask/target triplets	12,734 train / 685 test (2,358 rooms)
sceneforge-removal/sceneforge-removal-m/	SceneForge-Removal (M): multi-view triplets	3,803 train / 196 test (330 rooms)
SceneForge-IndoorDynamicEraser/	SceneForge-IndoorDynamicRemoval: all dynamic objects removed	2,313 train / 122 test
Removal-HardEffects/	60 real Pexels images with manual masks (qualitative benchmark)	60

SceneForge-16K = Removal-S train + Removal-M train = 16,537 triplets.

Manifests

Each JSONL row holds repo-relative paths (orig, mask, target). Removal-S images are split into shard sub-folders (000, 001, 002) to respect the Hub's per-folder file limit; always read paths from the JSONL. Train and test rooms are disjoint. Removal-S and Removal-M rooms are disjoint.

from datasets import load_dataset
ds = load_dataset("Jizhizi/SceneForge", "removal_s")
print(ds["train"][0])   # {'orig': ..., 'mask': ..., 'target': ..., 'scene_id': ..., 'mask_coverage_display': ...}

Reading the raw tars

SceneForge-S/ and SceneForge-M/ contain plain (uncompressed) tar shards plus an index.json that maps each room id to its shard. Paths inside a tar are SceneForge-S/<room_id>/... (or SceneForge-M/<room_id>/...).

import json, tarfile
from huggingface_hub import hf_hub_download

idx = json.load(open(hf_hub_download("Jizhizi/SceneForge", "SceneForge-S/index.json", repo_type="dataset")))
room = next(iter(idx["rooms"]))
shard = hf_hub_download("Jizhizi/SceneForge", f"SceneForge-S/{idx['rooms'][room]}", repo_type="dataset")
with tarfile.open(shard) as tf:
    members = [m for m in tf.getmembers() if m.name.startswith(f"SceneForge-S/{room}/")]

full_scene/ObjectIndexIDs_0000.exr holds per-pixel object ids (instance segmentation); object_index_map.json maps ids to object names. Read the EXR with OpenEXR or imageio.

Notes

• Removal-M camera types (from file names): base_camera 642, camera_similar 1,415, camera_diff 1,942. In the raw SceneForge-M/ tree the base camera folder is named based_camera.
• Raw SceneForge-S and SceneForge-M share 3 rooms; the Removal subsets do not.
• 48 rooms in SceneForge-S have no object-linked layers.
• Removal-HardEffects images come from Pexels and keep their original license; the rest is released under CC BY 4.0 (Infinigen assets: BSD-3-Clause).
• All content is synthetic; no personal data.

Citation

@misc{sceneforge2026,
  title         = {SceneForge: Structured World Supervision from 3D Interventions},
  author        = {Li, Jizhizi and Ao, Jiayang and Wicks, Danny and Tudosiu, Petru-Daniel},
  year          = {2026},
  eprint        = {2605.14399},
  archivePrefix = {arXiv}
}