Generative Cinematographer
Composing Camera and Object Motion in 3D
Project website · Paper · Code
Inference checkpoint for gencine, a geometry-conditioned video generation model. The provided examples generate 21 RGB frames at 832×480 with 50 denoising steps, CFG 3, sigma shift 5, BF16 precision and untiled decoding.
Files
model.json: model configuration and SHA-256 checksums.dit.safetensors: diffusion transformer.vace.safetensors: geometry-conditioning model.scale3d.safetensors: scale-conditioning model.vae_decoder.safetensors: video decoder.
Download
Install the Hugging Face CLI with pip install huggingface_hub, then download
the public checkpoint bundle:
hf download zonszer/Gencine_ckpts --local-dir weights/gencine
Inference
Install the gencine runtime following its project README. From the project root, run a provided example:
python -m gencine --weights weights/gencine \
--manifest examples/manifest.json --example hike --output outputs/hike_21
The runtime verifies weight checksums from model.json and runs offline with
preprocessed inputs. The examples and environment instructions are provided by
the gencine code repository.
See LICENSE and THIRD_PARTY_NOTICES.md for licensing.
Citation
If you use this project, please cite:
@misc{zhang2026generativecinematographer,
title = {Generative Cinematographer: Composing Camera and Object Motion in 3D},
author = {Zhang, Jiahan and Yang, Chaohao and Guruprasad, Namitha and Banerjee, Vivekjyoti
and Nguyen, Trong-Tung and Yuille, Alan and Bhattad, Anand},
year = {2026},
eprint = {2610.02180},
archivePrefix = {arXiv}
}