FilmTogether

Cinematic Video Generation from Low-Cost Inputs

Shiying ZhangXiaobin HuQianchi ZhangJiaqi ZhaoXiaoda YangBo YinShuicheng Yan

National University of Singapore

Read the paperCode on GitHubCineCraft-120K dataset

Real-world demos

Everyday footage. Cinematic worlds.

Seven everyday moments, transformed by FilmTogether. Original inputs on the left; actual model outputs on the right.

Basketball, reimagined as wuxia

1 / 7
Low-cost input
FilmTogether output
0:00 / 0:05
Read the editing prompt

Open output video

Synchronized, muted web previews with full clip durations. Real-world demonstrations are separate from the benchmark below.

CineCraft-Bench · Selected results

Low-cost footage.
Cinematic possibilities.

Ten curated benchmark examples. Compare FilmTogether outputs or switch to the paired high-cost dataset references.

Full-duration model outputs · exp5 step 6000 · 20 denoising steps

Imperial costume & jewelry

1 / 10
0:00 / 0:00
Read the editing prompt

Selected results

10 transformations, chosen for the showcase.

Preparing examples…

Muted, compressed web previews. Input and output clips retain their original duration. Dataset targets are reference high-cost videos, not model predictions.

The research

A production upgrade,
not a new performance.

Casual captures and professional productions can share the same scene, yet differ in lighting, clothing, materials, and visual atmosphere. FilmTogether studies how to improve these production elements without changing the underlying identity, composition, and motion.

The project brings together CineCraft-120K, a paired cinematic editing dataset, and a dual-memory adapter that enriches textual instructions and refines source-video hints.

CineCraft-120K

Paired scenes.
Element-wise instructions.

A cinematic source bank, structure-preserving counterpart synthesis, and prompts grounded in the actual low-to-high transformation.

120K+paired clips described in the paper
5broad cinematic domains
23fine-grained categories
17selected video examples on this page
CineCraft-120K data pipeline: source footage, quality curation, low-cost counterpart synthesis, and pair-aware instructions
The CineCraft-120K construction pipeline. Open the figure to inspect the details.
Dataset release in progress.

The public dataset repository is the release entry point. The videos above are a curated qualitative showcase, not an exhaustive evaluation. The full paired-video release is being assembled in batches.

Visit the dataset repository

FilmTogether

Two memories.
One coordinated transformation.

Expand what an instruction can express. Refine what source-video features pass to the generator.

FilmTogether architecture with instruction-routed text memory and layer-wise hint memory before VACE–DiT fusion
Instruction-routed Text Memory and layer-wise Hint Memory in the video editing pipeline.
01 / Semantic conditioning

Instruction-routed Text Memory

Learnable memory codes extend the instruction representation, providing additional capacity for coordinated appearance edits.

02 / Visual conditioning

Layer-wise Hint Memory

Prompt-conditioned, layer-specific corrections refine the source-video hints immediately before visual-context injection.

Explore and build

Paper, code, and data.

01 / Paper

Read the full manuscript

Method, data construction, experiments, and supplementary analysis.

Open PDF02 / Code

FilmTogether on GitHub

Public source, project-page assets, and release documentation.

Explore the repository03 / Dataset

CineCraft-120K

The public entry point for paired videos and editing prompts.

Visit Hugging Face

Citation

@misc{zhang2026filmtogether,
  title={FilmTogether: Cinematic Video Generation from Low-Cost Inputs},
  author={Shiying Zhang and Xiaobin Hu and Qianchi Zhang and Jiaqi Zhao
          and Xiaoda Yang and Bo Yin and Shuicheng Yan},
  year={2026},
  url={https://shiying-zhang.github.io/FilmTogether/}
}