linoyts/repo-to-space-example-videos

Gradio Space Example Inputs — Videos

A small, curated, freely-licensed pool of videos used as gr.Examples for Gradio Spaces that wrap video-input generation models (image-to-video, video-to-video, motion controls, etc.). Sister dataset for images: linoyts/repo-to-space-example-inputs.

When a Space takes video input, the agent building the Space picks 2–3 clips whose caption + categories match the model's task, downloads them via hf_hub_download, runs any model-specific preprocessing (trim / resize to expected shapes), and wires them into gr.Examples. The Space ships the preprocessed copies; this dataset is the source of truth.

The set is intentionally diverse on subject, motion, framing, and style so a caller can pick examples that flatter a specific model without needing to source new media each time.

Schema

metadata.jsonl — one JSON object per line, one record per asset:

{
  "file_name": "podcaster.mp4",
  "type": "video",
  "width": 720,
  "height": 1280,
  "duration_s": 9.6,
  "fps": 25.0,
  "categories": ["motion", "person", "indoor"],
  "caption": "A young woman with long lavender-purple hair ...",
  "source": "pexels-or-similar",
  "license": "free-to-use (CC0-style)"
}

The asset key is file_name (with underscore) — required by HF's folder_based_builder for the dataset viewer; other spellings (filename, path, ...) break it with SplitsNotFoundError.

Field	Type	Notes
file_name	string	Filename in this repo root.
type	"video"
width, height	int	Stored dimensions (post the source-shape constraints below).
duration_s, fps	float	Clip length and frame rate.
categories	list of strings	Coarse subject tags for fast filtering.
caption	string	Natural-language description: subject(s), setting/background, lighting, composition, motion, camera. No use-case judgments — the caller infers fit from the description.
source, license	string	Provenance / licensing.

categories are for cheap pre-filtering (drop dialogue for a pose-driven motion model, etc.). caption is for the finer matching step against the target model's task or example prompts.

Picking examples for a Space

1. Soft filter by categories — drop tags that are clearly off-task for the target model.
2. Rank by caption fit — read the surviving captions against the model's task description, trigger words, or example prompts; pick 2–3 the model will plausibly produce a good output on, not just any input that doesn't crash.
3. Diversify the final picks — different subjects, framings, or motion types — so the Examples row teaches users the breadth of what the Space handles.

from huggingface_hub import hf_hub_download
import json

DATASET = "linoyts/repo-to-space-example-videos"
meta = [json.loads(l) for l in open(
    hf_hub_download(DATASET, filename="metadata.jsonl", repo_type="dataset"))]

# ... filter + rank using `categories` and `caption` ...

chosen = ["podcaster.mp4", "man_dancing.mp4", "cat.mp4"]
paths = [hf_hub_download(DATASET, filename=f, repo_type="dataset") for f in chosen]
# preprocess paths (e.g. LTX wants 8k+1 frames, dims divisible by 32) and pass
# to gr.Examples in the Space.

Pre-bake the preprocessed copies into the Space repo. Don't rely on the dataset at Space runtime, and don't set cache_examples=True on ZeroGPU.

Source-shape constraints

• Videos: max 720 px on the short edge, ≤10 s, libx264 CRF 23, no audio.

Callers resize / trim further per model (e.g. LTX wants 8k+1 frames and dimensions divisible by 32).

License

All assets are sourced from royalty-free providers (Pexels / Unsplash and similar) under licenses that permit redistribution and modification without attribution. Released here as CC0-1.0 for simplicity.