README.md

metadata

title: Movetoy
emoji: 🟣
colorFrom: purple
colorTo: red
sdk: gradio
app_file: app.py
pinned: false
license: mit
short_description: Tiny from-scratch text-to-video toy (51M params), ZeroGPU
models:
  - Kompaktion/Movetoy

Movetoy

A 51M-parameter text-to-video model trained from random weights on one RTX 4070. Type anything, get a 2-second 32×32 clip of shapes doing (roughly) what you said. Merges and colour mixing are its party trick.

Model weights and full model card: Kompaktion/Movetoy.

Free API

No key needed. Arguments: prompt, seed (-1 = random), steps (8–64), guidance (1–6), seconds (2–15). Up to 2 s is one native pass; longer videos are stitched from 2-second windows.

Python:

from gradio_client import Client
video, info = Client("Kompaktion/Movetoy").predict(
    "a red circle and blue circle merge into a purple circle", -1, 32, 3.0, 2, api_name="/generate")
print(video, info)   # path to the downloaded .mp4, and "seed … · 16 frames (2.0 s)"

curl:

EVENT=$(curl -s -X POST https://kompaktion-movetoy.hf.space/gradio_api/call/generate \
  -H "Content-Type: application/json" \
  -d '{"data": ["a pink heart spins in the night sky", -1, 32, 3.0, 2]}' | jq -r .event_id)
curl -N https://kompaktion-movetoy.hf.space/gradio_api/call/generate/$EVENT

Runs on ZeroGPU; requests are queued, and heavy use draws on the caller's ZeroGPU quota.

How much training for such a simple task?

~382,000 procedurally generated captioned clips (23,874 steps × batch 16, every clip unique) — about 212 hours of 32×32 video, ~112 minutes on one RTX 4070 — to learn coloured shapes moving and merging. It still gets motion direction wrong about half the time. Details: model card.