Text-to-video model from scratch (2 brothers, 2 years, 2B params)

▶ See it at the source

What they did

Two brothers trained a 2B-parameter text-to-video model from scratch. Their write-up shows both good and bad samples.

How it went

The authors say it took 2 years. The HN post drew 158 points.

Worth knowing

• Price: unknown
• Status: sourced. We have not run this ourselves yet (no "Tested" badge).
• Quality bar (computed by code): 4/5 (primary source ✓, reproducible ✗, new ✓, example output ✓, safe ✓)

Related

model▶ YouTube demo

Veo (Google DeepMind)

Google DeepMind's page for the Veo video model family, with capability notes and sample clips straight from the lab.

Demo by Google DeepMind

Veo✓ vendor-checked

model▶ demo at source

Veo 3 on fal

Veo 3 served through fal's API, with text-to-video and optional generated audio. Price depends on whether audio is on.

Demo by fal · example via fal

Veofal✓ vendor-checked

model▶ demo at source

Veo 3 Fast on fal

A faster, cheaper Veo 3 variant on fal. It's a sensible default for drafts before you spend on the full model.

Demo by fal · example via fal

Veofal✓ vendor-checked

model▶ YouTube demo

Veo on Gemini API docs

Google's developer docs for generating video through the Gemini API, including when to pick Veo 3.1 versus faster editing-oriented models.

Demo by Google DeepMind

Veo✓ vendor-checked

tool▶ demo on X

Google Flow

Google's filmmaking app built on Veo and its image models. You assemble shots and scenes in one place instead of generating clips one by one.

Demo by @GoogleLabs

Veo✓ vendor-checked

tool▶ demo at source

Replicate video collection

Replicate's curated collection of text-to-video models you can run by API, an alternative host to fal.

Demo by Replicate

ReplicateVeoHailuo / MiniMax✓ vendor-checked