V-RAE¶

Development line: project:v-rae · thread v-rae-development
Last event: 2026-08-24 · 1 dated since 2026-08-24 · Researched: 2026-09-05 · confidence: high

What it is¶

V-RAE is a research codebase and checkpoint set for video reconstruction and generation. It pairs frozen DINOv3, SigLIP2, V-JEPA2.1, or EUPE encoders with temporal pooling and a video decoder.

• Video reconstruction from frozen representations.
• Class-conditional generation across video benchmarks.
• Future-frame prediction without updating encoder weights.

Best reported generation results are 117.86 gFVD on UCF101 and 19.16 on Kinetics-600. The code is research-oriented and requires Linux, NVIDIA CUDA, FFmpeg, Python 3.10+, and matching encoder weights.

Development line¶

• 2026-08-24 — V-RAE project resources were linked. On 2026-08-24, the project published entry links to its website, GitHub repository, and Hugging Face model page without separate release details or verified capabilities.

What changed¶

• 2026-08-13 — The V-RAE paper was submitted to arXiv, defining the representation-autoencoder approach and tFVD temporal-coherence diagnostic.
• 2026-08-14 — The official implementation began with an initial GitHub release.
• 2026-08-19 — The repository added data manifests, visual results, and a fix for mismatched configurations.
• 2026-08-24 — Public entrypoints covered the project page, official implementation, and checkpoint repository; no first-party source proves a distinct code or model release on this date.
• 2026-08-26 — The repository updated GIF visual results.

How to use this¶

As of 2026-08-24, treat the linked V-RAE website, GitHub repository, and Hugging Face model page as starting points to check project documentation, source, and model availability; exact versions and usage requirements still need research.

1. Clone the implementation, create a Python 3.10 environment, install FFmpeg and the pinned package dependencies. — https://github.com/V-RAE/V-RAE
2. Download the V-RAE checkpoint repository into ckpts, then download the frozen encoder weights that match the selected variant. — https://huggingface.co/Guomh0707/V-RAE-Models
3. Put input clips in assets/sample1.mp4 through assets/sample3.mp4, then run python sampling.py dino, siglip, vjepa, or eupe; compare results under outputs/<variant>/. — https://github.com/V-RAE/V-RAE

Best practices¶

• Match every V-RAE checkpoint to its corresponding frozen encoder weights; the checkpoints are custom PyTorch files intended for the official codebase. — https://huggingface.co/Guomh0707/V-RAE-Models
• Use Linux with an NVIDIA GPU, a CUDA-compatible driver, FFmpeg, and Python 3.10 or newer; verify CUDA and TorchCodec before inference. — https://github.com/V-RAE/V-RAE
• Treat the published numbers as task-specific research results: rFVD measures reconstruction, gFVD generation, and tFVD temporal coherence; do not compare them as a single universal quality score. — https://github.com/V-RAE/V-RAE

Superseded by this¶

• 2026-08-19 — Initial repository configurations were superseded by the documented fix for mismatched configs.

Still unknown¶

• No dated first-party release record exists for 2026-08-24 itself; the linked paper and initial repository release predate that date.
• The model hub omits a dated revision on the retrieved page, leaving the original checkpoint-publication date unverified.
• No hosted inference endpoint or production deployment guidance exists; the model card says no Inference Provider deploys it.

Sources¶

Agent brief¶

• Subject: project:v-rae, thread v-rae-development, 1 dated events 2026-08-24 → 2026-08-24.
• Practical note: As of 2026-08-24, practitioners should treat the linked V-RAE website, GitHub repository, and Hugging Face model page as the starting points for checking project documentation, source, and model availability; exact versions and usage requirements still need research.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.