Vidi

Why it's ranked here

The project pairs usable model code with unusually broad evaluation tooling. It covers temporal retrieval, object grounding across time, plot reasoning, character tracking, and question answering. The strongest practical evidence is the released benchmark data, evaluators, model weights, inference code, and finetuning support. However, the repository presents a research workflow more clearly than a production integration surface.

What's good

The evaluation support goes beyond a single headline metric. Temporal retrieval reports overlap, precision, and recall across query length, video duration, and audio or visual dependence. Spatial grounding measures both frame presence and bounding-box volume over time. Plot evaluation separately covers exact-answer questions and character segments, including transcript error and box overlap. This makes model comparisons inspectable instead of opaque.

Tradeoffs

The hosted demo recommends English queries, modest video resolution, and patience because long videos can take several minutes. Local inference expects CUDA, half-precision tensors, Flash Attention, FFmpeg, and FFprobe, so setup is not lightweight. The published model inference centers on temporal retrieval, while the wider demo capabilities are not all exposed through an equivalent documented local interface. One evaluation path can enter an interactive debugger after malformed results.

How to use it well

Vidi fits researchers and video teams testing search, grounding, chaptering, highlights, question answering, or automatic rough cuts. Start with the hosted demo to validate a task, then use the released weights and command-line evaluators for repeatable experiments. Keep videos moderate in resolution and format predictions exactly as documented. It does not supply a documented production API, managed deployment layer, or lightweight CPU workflow.

Observed

Primary language

Python

Interfaces

Hosted web demo plus command-line inference and evaluation scripts

Model distribution

Vidi-7B and Vidi1.5-9B weights are linked through Hugging Face

Local inference requirements

CUDA execution, half-precision tensors, FFmpeg, and FFprobe

Evaluation coverage

Temporal retrieval, spatio-temporal grounding, plot reasoning, character evaluation, and video question answering

Data formats

Evaluation inputs and outputs use CSV, JSON, and JSONL structures