VideoCrafter2
Tencent's data-efficient open video diffusion model that decoupled motion from appearance instead of demanding massive video datasets.
About
VideoCrafter2 ('Overcoming Data Limitations for High-Quality Video Diffusion Models', January 2024) came from Tencent AI Lab at a time when every serious video model seemed to need an internet-scale video corpus. Its trick was to decouple the problem: train the motion module on abundant low-quality video while fine-tuning the spatial layers on high-quality synthesized images, then fully train the coupled spatial-temporal modules together so motion stays natural and text alignment stays robust. The result was a compact open model for text-to-video and image-to-video generation that punched above its data budget.
It has since been eclipsed by larger open models like HunyuanVideo and the Wan family, but it remains a useful lightweight baseline with 5.1k GitHub stars and an actively maintained repo. One caveat for reusers: the GitHub repo carries no machine-readable license (GitHub reports NOASSERTION), even though the code is distributed as an open research release, so commercial users should confirm terms with the authors rather than assume Apache-2.0.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published benchmark scores; data-efficiency claims are author-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Tencent AI Lab paper arXiv 2401.09047 (Jan 2024); GitHub AILab-CVC/VideoCrafter (5.1k stars, updated Oct 2026). Strengths: influential data-efficiency approach, long maintenance history, real community use. Docked for: (1) no machine-readable license on the repo (NOASSERTION via GitHub API) despite open distribution, a commercial-use risk; (2) superseded by larger open models, no published benchmark scores. License: undetermined from repo metadata; flagged for confirmation.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Tencent AI Lab; paper arXiv 2401.09047 (Jan 2024); 5.1k stars
- · Data-efficient: motion module on low-quality video + spatial layers on synthesized images
- · Superseded by larger open models; no published benchmark scores