Research map / Video / Video generation
Captain Cinema: Towards Short Movie Generation
Junfei Xiao, Ceyuan Yang, Lvmin Zhang, Shengqu Cai, Yang Zhao, Yuwei Guo, Gordon Wetzstein, Maneesh Agrawala, Alan Yuille, Lu Jiang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GoldenMem compressed context at 16 pairs (ablation) | Captain Cinema keyframe generator · GoldenMem ablation, 16 pairs | Compute (PFLOPS est.): 30 -> 21 | Improved |
| add GoldenMem at 16 pairs (ablation) | Captain Cinema keyframe generator · GoldenMem ablation, 16 pairs | Visual consistency: 4.7 -> 4.3 | Hurt |
| add GoldenMem at 32 pairs (ablation) | Captain Cinema keyframe generator · GoldenMem ablation, 32 pairs | Compute (PFLOPS est.): 55 -> 35 | Improved |
| add GoldenMem to enable 48 pairs (ablation) | Captain Cinema keyframe generator · GoldenMem ablation, 48 pairs | feasibility: OOM -> 52 PFLOPS | Improved |
| add multi-frame interleaved finetuning of video model (ablation) | Captain Cinema · Long-video eval (VBench-2.0 style) | Temporal dynamic / aesthetic / quality: Aesthetic 56.8 -> 57.2; Quality 60.9 -> 61.7; Dynamic 64.4 -> 65.4 | Improved |
| add multi-frame interleaved finetuning of video model (ablation) | Captain Cinema · Long-video eval (VBench-2.0 style) | Temporal consistency: 91.9 -> 91.0 | Hurt |
| use Captain Cinema instead of LCT in long contexts (24 pairs) | interleaved long-context generation · Stress test, 24 context pairs | Identity consistency: 0.14 -> 0.42 | Improved |
| use Captain Cinema instead of LCT for temporal consistency | long video generation · Long-video eval (VBench-2.0 style) | Temporal consistency: 94.8 -> 91.0 | Hurt |
| use Captain Cinema instead of LCT | long video generation · User study | video quality: 2.4 -> 3.3 | Improved |
| progressive long-context finetuning with modest warmup (step 40,000) instead of direct finetuning or late-checkpoint start (ablation) | keyframe generator (Flux 1.Dev) · keyframe generation | training stability | Improved |
Closest papers by meaning
- MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequences (ICLR 2025 · similarity 0.92)
- MovieBench: A Hierarchical Movie Level Dataset for Long Video Generation (CVPR 2025 · similarity 0.90)
- SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction (ICLR 2024 · similarity 0.89)
- CineTechBench: A Benchmark for Cinematographic Technique Understanding and Generation (NeurIPS 2025 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.88)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
