README.md

metadata

license: apache-2.0
task_categories:
  - video-text-to-text
  - question-answering
language:
  - en
tags:
  - world-model
  - spatial-intelligence
  - seva
  - qwen3-vl
  - vsi-bench
  - ovos-bench
  - mmsi-bench
pretty_name: >-
  WorldFold v7-enhance++ Checkpoint & Precomputed 4-Anchor SEVA World-Model
  Feature Caches

WorldFold (v7-enhance++) — Model Weights & Unified 4-Anchor SEVA World-Model Feature Caches

This repository hosts the trained Worldfold-v7-enhance++ adapter checkpoint and the precomputed Stable Virtual Camera (SEVA v1.1) + Qwen3-VL-8B Visual Tower world-model feature caches across all four training and evaluation splits (VSI-Train-10k, VSI-Bench, OVO-S-Bench, and MMSI-Video-Bench).

All four splits are 100% unified under the 4-Anchor Chronological Protocol (four_anchor_protocol: True):

• 4 Chronological Anchors (M = 4) sampled at (20%, 40%, 60%, 80%) of the video timeline.
• 1 Canonical 3D Camera Trajectory per Anchor × 2 Rollout Steps (T = 2) (Orbit-Right, Orbit-Left, Dolly-Forward, Pan-Sweep).
• 8 Generated Novel Views (4 × 2), each encoded by frozen Qwen3-VL-8B-Instruct visual tower and pooled to 16 patches ([4, 2, 16, 4096], bfloat16, L2 norm 12.32).

Codebase branch: inFaaa/WorldFold (v7-enhance-pp)

---

Included Assets

File in Repository	Split / Role	Count	Protocol (M × T × P × D)	Size	Unpack / Target Path
checkpoints/Worldfold-v7-enhance++.pt	2-Epoch Trained Adapter Checkpoint (fp32)	1 checkpoint	K in {256, 128, 64, 16}	75.5 MB	outputs/nested_prior_runs/spec_v7_runs/checkpoints/Worldfold-v7-enhance++.pt
vsi_eval_seva_288scenes.tar.gz	VSI-Bench Test SEVA Features	288 scenes (5,130 QA)	4 Anchors × 2 Steps × 16 × 4096 (four_anchor_protocol=True)	549.3 MB	Extract into outputs/cache/vsi_eval_cache/ (scene_feats_real_seva/*.pt)
ovos_eval_seva_339videos.tar.gz	OVO-S-Bench Test SEVA Features	339 videos (1,695 QA)	4 Anchors × 2 Steps × 16 × 4096 (four_anchor_protocol=True)	551.6 MB	Extract into outputs/cache/ovos_eval_cache/ (real_seva_feats/*.pt)
mmsi_eval_seva_1106questions.tar.gz	MMSI-Video-Bench Test SEVA Features	1,106 questions	4 Anchors × 2 Steps × 16 × 4096 (four_anchor_protocol=True)	1.84 GB	Extract into outputs/cache/mmsi_eval_cache/ (real_seva_feats/question_*.pt)
vsi_train10k_seva_only_3449scenes.pt	VSI-Train-10k Training SEVA Features	3,449 scenes (10,169 QA)	4 Anchors × 2 Steps × 16 × 4096 (four_anchor_protocol=True)	8.26 GB	Packed dict {scene_key: {wm_4x2, real_4x2, anchor_context, real_context, metadata}}

---

Quick Download & Unpack

pip install -U huggingface_hub

# Download all checkpoints and SEVA caches
hf download Jinfa/WorldFold-v7-Cache --repo-type dataset --local-dir ./WorldFold-v7-Cache

# Unpack the 3 evaluation benchmark SEVA caches
mkdir -p outputs/cache/vsi_eval_cache outputs/cache/ovos_eval_cache outputs/cache/mmsi_eval_cache
tar -xzf ./WorldFold-v7-Cache/vsi_eval_seva_288scenes.tar.gz -C outputs/cache/vsi_eval_cache/
tar -xzf ./WorldFold-v7-Cache/ovos_eval_seva_339videos.tar.gz -C outputs/cache/ovos_eval_cache/
tar -xzf ./WorldFold-v7-Cache/mmsi_eval_seva_1106questions.tar.gz -C outputs/cache/mmsi_eval_cache/