Skip to content
← Models
Open weightsShanghai AI Laboratory/10 trustMetrics-checked

Vchitect-2.0

A parallel-transformer text-to-video research model from Shanghai AI Lab's Vchitect team, with open checkpoints and 2K upscaling via VEnhancer.

Visit official page →

About

Vchitect-2.0 ('Parallel Transformer for Scaling Up Video Diffusion Models', January 2025) comes from the Vchitect team at Shanghai AI Laboratory, the same group behind the VBench video evaluation suite. The paper's focus is systems engineering for video diffusion: a hybrid parallelism framework mixing data and sequence parallelism with recomputation and offloading, so long, high-resolution videos can actually be trained on distributed hardware. The base text-to-video model generates up to 720x480 at 8fps, and the pipeline pairs it with VEnhancer to upscale to 2K and interpolate to 24fps.

Inference code and checkpoints were released in September 2024 under Apache-2.0, explicitly allowing free commercial use, and the team later open-sourced the Vchitect-T2V-Dataverse training dataset. It is a smaller community effort (around 900 stars) with no widely cited benchmark scores of its own, but its provenance from the VBench authors gives its evaluation claims more weight than the average research release.

Specs

Context window

40 frames default at 768x432 (up to 720x480), 8fps; VEnhancer upscales to 2K 24fps

Pricing

Free and open source (Apache-2.0); self-hosted

Free quota

Full access, no quota, self-hosted

API access

No official API; inference code and checkpoints on GitHub and Hugging Face

Weights

Open weights

Released

Not disclosed

Benchmarks

Author-reported benchmark evaluations in the paper; no widely cited independent scores

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-video researchdistributed training techniquesvideo upscaling pipelinebenchmark-aligned evaluation

Trust score — 6/10

/10 trustMetrics-checked

Scored 6/10 (metrics tier, metrics-v1). Evidence: Vchitect team (Shanghai AI Lab) paper arXiv 2501.08453 (Jan 2025); GitHub Vchitect/Vchitect-2.0 (921 stars, Apache-2.0 confirmed, checkpoint released Sept 2024 per repo README); Vchitect-T2V-Dataverse released Mar 2025. Strengths: permissive license incl. commercial use, open dataset, credible eval provenance (VBench authors). Docked for: (1) small community and no independent benchmark scores; (2) research-grade code with modest maintenance. License: Apache-2.0, confirmed on repo README and via GitHub API.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
5.0/10
Maintenance · recency, releases, issue load
6.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
6.0/10
  • · Vchitect team, Shanghai AI Lab (VBench authors); paper arXiv 2501.08453 (Jan 2025)
  • · 921 stars; inference code + checkpoint released Sep 2024
  • · Hybrid data/sequence parallelism framework; VEnhancer upscales to 2K/24fps
  • · Small community; no widely cited independent scores

Sources