Tutorials Video

Generating Video with AI: What the Tools State and What to Check

By Infinite Creation · Intermediate · ~14 min

A 1960s mid-century comic book style editorial illustration showing a director at a monitor with a generated mountain valley clip, storyboard cards behind and a clapperboard in front, while a colleague steps through the clip frame by frame on a jog wheel.

Overview

AI video generators create short clips from a text prompt or a still image. Some now generate matching sound. This guide is for creators who want to use clips as establishing shots, backgrounds or stylised scenes, and who need to know what the main tools state about themselves before they commit. It separates the hosted services from the open model you run yourself, lists the questions about licences and labelling to settle first, and explains why a generated clip needs a closer inspection than footage from a camera. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, and the Sources section says what each page established.

Kit and Prerequisites

A clear idea of the shot, in one sentence, before you open a tool
The tool's terms of use, read for commercial use and ownership of output
For the open model, a computer with a capable graphics card and ComfyUI
A frame-by-frame review of every generated clip before it enters an edit

Working Through It

Pick the shot type the tools do well

Generators handle short, self-contained shots: an establishing view, abstract motion, a stylised scene. They struggle with the same character across shots, readable text and physical accuracy. Plan the edit so no generated shot has to match another exactly.

Understand the three kinds of tool

Hosted consumer services such as Kling, whose page names Kling 4.0. Hosted platforms with wider editing such as Runway, whose page calls Runway Creative a cloud-based platform and names Gen-4.5. Models from large companies such as Google's Veo, whose DeepMind page names Veo 3.1. Each is accessed through its maker's service.

Read what the page says about sound and watermarks

Google's Veo page states Veo generates sound effects, ambient noise and dialogue with the video, adds that natural spoken audio for short segments is still an area of active development, and states output is watermarked with SynthID, Google's technology for detecting AI-generated content. It lists access through the Gemini app, Flow and the Gemini API.

Consider the open model if you want local control

Wan 2.2 is released under the Apache 2.0 licence, which the ComfyUI announcement says permits commercial use. The announcement lists Wan2.2-TI2V-5B, Wan2.2-I2V-A14B and Wan2.2-T2V-A14B, and says the 5B version has optimised memory use. It gave no memory figure, so confirm the requirement in ComfyUI's documentation before downloading.

Check licences and training data claims

Moonvalley's page describes its Marey model as trained on licensed data and as commercially safe, without naming the data and without an indemnity statement. A trade report we did not read in full says indemnity covers its data contractors only. Kling's page states no commercial-use terms. Read the terms of every tool before commercial use.

Review, label and keep the prompt

Step through each clip frame by frame for warped hands, melting objects, changing faces and invented text. Label the clip as AI-generated where your audience would care. Save the prompt, model version and date with the file.

Worth knowing:

Generate several versions of a shot. Differences between them show what is invented.
Do not generate a depiction of a real person. It is a fabrication and carries legal and ethical risk.
Models change quickly. Note the version you used, because the same prompt gives different results later.

Knowledge Base

Reference Notes

The decisions that matter are about rights and honesty more than about prompts.

Hosted services versus a model you run

A hosted service handles the hardware and updates the model without notice, which is convenient and means the tool you tested last month is not necessarily the tool you use today. A model such as Wan 2.2 is pinned to a version and run on your machine, which gives repeatability and privacy, and demands a graphics card and setup time. Either is a sound choice. What decides it is whether you need repeatable results or the fastest start.

Legal risk is unevenly settled

Training data disputes and the status of generated output differ between countries and between tools. A tool that states it uses licensed data reduces one risk and does not necessarily remove it, and we did not read the underlying agreements. If a client or broadcaster will scrutinise your work, ask what they require before you generate anything for it.

Where This Fits

A generated clip replaces a shot, and Video Production Fundamentals covers how planning, camera, sound and edit constrain every shot in a finished video.

FAQ

Q: Which generator is best?
A: We did not test any, so we cannot rank them. The makers' pages above state different strengths.

Q: Is selling a video made with these tools allowed?
A: It depends on the tool's terms and plan. Kling's page states no terms and Wan 2.2's announcement states Apache 2.0 with commercial use. Read each tool's terms.

Q: Do I have to label generated video?
A: Platform rules and some laws increasingly require it. See the labelling page and check the rules where you publish.

Vocabulary

Text to video Generating a video clip from a written description. Image to video starts from a still picture and animates it. Both produce short clips and neither records anything that happened.

Watermark A mark added to generated output to show its origin. SynthID, Google's system, embeds a signal that detectors read. Platforms and editing often remove visible marks, and invisible ones vary in durability.

Indemnity A legal promise by a company to cover costs if a customer is sued over what the company supplied. Whether a generator offers one, and for whom, is stated in its contract and not on its marketing page.

Sources

Maker pages were read on 8 October 2026. No clip was generated. Checked 8 October 2026.

Google DeepMind, Veo. States Veo 3.1 is current, generates sound with video, watermarks output with SynthID, and is available through the Gemini app, Flow and the Gemini API.
Kling. Names Kling 4.0 for video and Kling Image 3.0. States no free credits and no commercial-use terms.
Runway. Describes Runway Creative as a cloud-based platform, names Gen-4.5, and shows a try-free offer without terms.
ComfyUI, Wan 2.2 support. States Apache 2.0 licensing permitting commercial use, the three model variants, and gives no memory figures.
Moonvalley, Marey. States the model is trained on licensed data and is commercially safe, names no data sources, and makes no indemnity statement.

The trade report on Moonvalley's indemnity was not read in full. No page states quality results, maximum clip length or cost terms that we relied on. Model names and versions change often, so check each page for the current state.