anonymous111111/generative-world-renderer-clips

A Scaling Recipe for Generative World Renderer

NeurIPS 2026 Evaluations & Datasets Track — anonymous submission.

Reviewer Sample — NeurIPS 2026 Submission. This repository currently hosts a 40-clip flattened reviewer sample (≈ 2.8 GB) so that NeurIPS 2026 reviewers can inspect data quality and format without per-request gated access. The full dataset described in the accompanying paper — approximately 4 M frames, 40 hours of playtime, 720p / 30 FPS, ~11 k sub-clips, with five synchronized G-buffer channels (depth, normals, albedo, metallic, roughness) and a parallel motion-blurred RGB variant — totals roughly 2 TB and therefore exceeds the 300 GB single-repository limit for public Hugging Face datasets. The complete release will be made publicly available via gated access by the NeurIPS 2026 camera-ready deadline, in compliance with the Datasets & Benchmarks public-release requirement. Anonymous access links to the full release will be provided in the supplementary material of the submission. The schema, license (CC BY-NC-SA 4.0), and Responsible-AI metadata (provided in the accompanying croissant.json at the repository root) apply identically to both the reviewer sample and the full release.

Schema (sample config)

#	Column	Type	Description
1	game	string	Source title: cyberpunk2077 or black_myth_wukong.
2	clip_id	string	Stable identifier for the sub-clip.
3	rgb_path	string	Repo-relative path to the RGB video file.
4	depth_path	string	Repo-relative path to the depth G-buffer video file.
5	normal_path	string	Repo-relative path to the camera-space normals G-buffer video file.
6	albedo_path	string	Repo-relative path to the albedo / base-color G-buffer video file.
7	metallic_path	string	Repo-relative path to the metallic G-buffer video file.
8	roughness_path	string	Repo-relative path to the roughness G-buffer video file.

Intended Use Cases

• Video inverse rendering (depth / normals / albedo / metallic / roughness from RGB).
• G-buffer-conditioned forward generative rendering and neural relighting.
• Controllable video editing of AAA-style game footage (style, lighting, weather, visual-effect transfer).
• Material-decomposition and intrinsic-image benchmarks on long, dynamic, in-the-wild-style sequences.
• Temporal-consistency research for video diffusion, video depth, and video normal estimators.
• Synthetic-to-real transfer and motion-blur robustness studies (paired sharp / blurred RGB).
• Development and benchmarking of VLM-as-judge protocols for material-channel quality assessment.

License

The dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0), in accordance with each source game developer's Fan Content Policy and EULA, which permit non-commercial derivative works and sharing.

The full release is gated: researchers will be required to sign a Terms-of-Use agreement acknowledging the source-game copyrights and committing to non-commercial research use before access is granted. The toolkit used to curate the dataset will be open-sourced separately to enable reproducible extension to additional games.

Limitations

• Domain coverage is biased toward the two source titles' art directions; under-represents medical, microscopy, and aerial photometric phenomena.
• G-buffer extraction relies on offline RenderDoc inspection and per-title ReShade hooks; porting to a new game requires non-trivial engineering effort.
• Camera-space normals are reconstructed from the depth buffer via inverse projection and finite differences, which can introduce high-frequency noise on thin or aliased structures.
• The released VLM-based evaluation protocol is a complementary, not a replacement, signal to ground-truth metrics; it inherits the judge model's failure modes on ambiguous coated / painted metals, wet surfaces, translucency, and compression artifacts.

Ethics, Safeguards, and Broader Impact

• No personal data. All visual content is synthetic game-engine output. Any depicted humanoid figures are fictional in-game characters; no real-world faces, voices, geolocation, medical, demographic, political, religious, or socio-economic personal data are collected, derived, or distributed.
• Misuse mitigation. Potential misuse — generating manipulated or stylized game footage without attribution — is mitigated by (i) gated access with a signed Terms-of-Use restricting use to non-commercial research, (ii) the CC BY-NC-SA 4.0 license that propagates non-commercial and share-alike obligations to derivative works, and (iii) preservation of the original games' watermarks / HUDs in every captured frame so downstream re-renders remain identifiable as game-derived content.
• User study. A 25-expert pairwise preference user study was conducted as part of the accompanying paper. Participation was voluntary and uncompensated; no personally identifying information was collected.