Qwen, Luma and Pika: testing AI video creation that’s hard to tell from real footage

1.5 years ago, AI videos easily gave themselves away: jerky movements, blurry faces, unnatural physics of objects. Now such videos have become noticeably less frequent in feeds — but there are two explanations for this phenomenon, and they are fundamentally different. Either the technology has advanced so much that AI videos have become difficult to distinguish from real ones. Or people have simply learned to use the tools more professionally. Most likely, both factors are working simultaneously.

What changed over 1.5 years

The quality of video generation has grown by leaps and bounds. If previously the very first frames immediately revealed the neural network origin of the video, modern tools have learned to imitate real cinematography: camera movement, depth of field, natural light transitions between scenes. At the same time, user skill has grown. The community around video generation has developed stable practices — how to compose prompts, what styles to set, how to post-process the result. What previously required dozens of iterations and produced mediocre results is now reproducible in a few attempts.

Three tools that currently set the tone in the market:

• Qwen Video from Alibaba — a multimodal model with emphasis on precise adherence to text instructions
• Luma Dream Machine — specializes in cinematic scenes with smooth camera movement
• Pika — tailored for short viral videos, popular with content creators on TikTok and Reels

What exactly was tested

The authors set a specific task: try to create with each tool a video that looks organic in a social media feed and doesn't trigger an instinctive "this is obviously AI". Not just technically beautiful imagery, but content with plausible dynamics, correct rhythm, and details that viewers perceive subconsciously.

"We decided to conduct an experiment and find out whether poor-quality

videos created by neural networks have really disappeared, or whether people have learned to make such good videos that we no longer notice who is who."

The task turned out to be more difficult than expected. Each of the three tools has clear strengths and weaknesses. Qwen works better with detailed prompts, but can produce artifacts when people move in the frame. Luma creates impressive atmospheric scenes, but struggles with close-ups of faces and dialogues. Pika works quickly and intuitively, but quality decreases with complex multi-layered scenes with many objects in the frame.

Where is the boundary between real and generated

Modern models have learned to hide classic signs of AI generation. Fingers no longer blur into incredible configurations, text on background objects has become readable, and the physics of liquids and fabrics has improved significantly. Nevertheless, several persistent markers remain noticeable to a trained eye:

• Lighting that is too "perfect", which you rarely see in real amateur filming
• Micro-movements of the eyes and facial expressions that don't quite match speech
• Hair textures when moving, especially with a complex or dynamic background
• Editing transitions and angles that are atypical for human operators

It is precisely in these details that the three tools diverge in results. Luma wins on landscapes and object shots where there are no close-ups of people. Pika holds up better where you need fast dynamic editing without lingering on details. Qwen shows itself where precision in reproducing small details of a scene by description is important.

What this means

The boundary between AI video and real filming is rapidly blurring. For content creators, this opens up new opportunities — to make videos without cameras, actors, and studios, lowering the barrier to entry to almost zero. For viewers, this raises a fundamentally new question: what should you now rely on when assessing the authenticity of video in your feed?