Kling 3.0 : Maîtrise l'IA Vidéo N°1 en 2026 (Tuto Complet)

Overview

Thomas presents Kling 3.0 as a video-generation model for creating coherent, multi-scene videos from prompts. He demonstrates its features and a workflow using fal.ai, which he prefers because it charges per use rather than requiring a monthly subscription. He also provides a free “metaprompt”—a prompt designed to turn a rough idea into detailed Kling instructions.

Kling 3.0 Features

1. Improved lip-sync and audio generation: Kling can generate speech and video together, with the speaker’s mouth movements matching the audio.
2. Multi-shot generation: Videos can be divided into separately described shots, each with its own duration. The shots are generated together to help maintain consistency in characters and scenes.
3. Videos up to 15 seconds: Thomas contrasts this with the shorter maximum durations of earlier models.
4. Elements: Users can upload reference images of people, characters, products, or other objects, then refer to them in prompts with an @ mention. Multiple images and angles can improve visual consistency.
5. Physics: Thomas reports progress in how objects move, but says his bouncing-basketball test still had problems, including the ball changing shape and direction.
6. Camera movement: He highlights precise camera instructions as a strength for creating cinematic-looking shots.

Tutorial Workflow and Practical Notes

Thomas uses Claude to help write prompts. His metaprompt asks follow-up questions about the video type, number of shots, elements, starting image, and scene details. He recommends generating prompts in English.

In fal.ai’s multi-shot workflow, leave the single-prompt field empty, add a prompt and duration for each shot, and keep the total video length within 15 seconds. Thomas notes a possible interface quirk: element references may not appear highlighted in multi-shot prompts, though he says they still work.

For more consistent characters, upload reference images from several angles. For fictional characters without suitable reference images, Thomas describes creating reference views with an image generator and cropping them into separate images.

Choose an aspect ratio suited to the intended platform, such as 16:9 for YouTube or 9:16 for vertical video. Audio can be disabled to reduce the generation cost.

A Voice ID option can help keep a voice consistent across scenes and generations. Thomas advises considering copyright when using someone else’s voice.

The demonstrations include a multi-scene sequence featuring Thomas and other referenced characters. He also notes that using a limited or profile-only reference image can result in an inconsistent face.

Applications Discussed

Thomas suggests using Kling to create attention-grabbing social videos and AI-generated advertisements. He describes an ad workflow that combines a consistent character, voice, setting, and product.

He cites a YouTube channel and examples of high-view videos as evidence of audience potential, and says his agency has produced AI content for clients. These are his examples and claims, not an independent assessment of performance or commercial results.

Main Speaker and Source

Thomas Berton, presenting as the head of the AI automation agency Azuro AI.