Haiper is the newest AI startup to introduce its personal generative video mannequin, becoming a member of firms like Pika Labs, Runway, and OpenAI’s Sora to deal with AI storytelling.
The corporate is concentrated on high quality and is presently producing brief two-second clips, nevertheless it has massive ambitions, together with aiming for Sora-like realism sooner or later.
Like OpenAI and Anthropic, Haiper is working towards synthetic basic intelligence (AGI), constructing what it calls a “highly effective perceptual basis mannequin.”
Within the clips I generated it appeared to have a extra native understanding of motion, not requiring in depth cueing or movement controls to get the wing flapping or dancing excellent. Haiper is also currently free.
What’s Haiper?
Haiper is the brainchild of former Google DeepMind and TikTok engineers and brings collectively cutting-edge machine studying abilities to democratize creativity.
I spoke to Dr. Yishu Miao, co-founder and CEO of Haiper concerning the product and to get a demo of the way it works, together with a batch era characteristic for creators.
I requested him what units it aside from the rising variety of AI video platforms and he stated they had been “hyper-focused on video,” creating spectacular generative movement with out requiring in depth, speedy engineering.
The objective, he advised me, is to create a easy product that anybody can use, specializing in getting the coaching information and labeling proper to make instructions simpler.
How properly does Haiper work?
One factor I observed about Haiper is how intuitive the underlying mannequin is for figuring out the proper movement for a video. The clips are brief and I have not been in a position to attempt to lengthen them, however for 4 seconds of video it captured very properly what I anticipated.
Dr. Miao stated it truly works higher with shorter prompts and with out utilizing the movement slider, because the AI mannequin anticipates motion higher than people.
In an indication of this functionality, he created a video of a forest at twilight with movement set to most, then one other with movement set to 1 (the default). When he went all out, he had issues not solely with coloration but additionally with inserting stars and bushes in locations the place they do not belong.
After altering the settings and leaving the motion as much as the AI, it captured the sky transferring within the background behind the bushes and even the rustling of the bushes themselves. She felt extra pure.
Testing Haiper with 7 clips
I got here up with seven video clip concepts designed to check each your realism and movement capabilities from a comparatively brief message. He did a surprisingly good job on most of them, however had issues with animals and other people.
a hummingbird
(Picture: © Haiper/AI video)
First I requested him to indicate “A hummingbird fluttering and feeding on a vibrant, nectar-rich flower in a lush backyard.” That is one thing different fashions can battle with, significantly across the wings, however Haiper appeared to do a very good job, with out going overboard an excessive amount of or fusing the wings collectively.
Dancing in a highlight
(Picture: © Haiper/AI video)
I then requested him to make me a video of “A talented dancer performing an intricate, fast-paced pirouette on a lighted stage.” Barely sluggish movement and whereas the primary two seconds had been slick, we blurred and blended within the second half.
dolphin soar
(Picture: © Haiper/AI video)
Then I challenged him to indicate “An impressive dolphin leaping out of the ocean, spinning within the air, and splashing again into the water.” He actually struggled right here. It seems to be cool nevertheless it’s not lifelike and the dolphin’s fin appeared to look out of nowhere.
Scorching air balloon
(Picture: © Haiper/AI video)
After I requested him to create a “colourful sizzling air balloon rapidly ascending into a transparent blue sky dotted with fluffy white clouds,” I had fairly excessive expectations, since this needs to be straightforward. He captured the clouds and sky, shaped a wonderful sizzling air balloon, however he did not get the momentum proper, because it had a small elevation.
Soccer participant
(Picture: © Haiper/AI video)
He spectacularly did not seize “knowledgeable soccer participant kicking a ball exhausting, sending it into the highest nook of the objective.” Haiper fused the participant with himself creating a wonderful however unusual mass of motion.
Fireworks within the sky
(Picture: © Haiper/AI video)
I believe it is more durable to do fireworks mistaken than to do them proper, so the thought was to see how properly they adopted instructions. I ordered it for “a mesmerizing show of fireworks that explode in speedy succession, illuminating the night time sky with sensible colours.” He did not do a foul job. Began off properly with the colours however took the explosion half too actually.
considerate robotic
(Picture: © Haiper/AI video)
I lastly examined its capacity to animate a nonetheless picture. I gave him an Adobe Firefly-generated picture of a robotic trying into area with the textual content message “pensive robotic watches the sky transfer.” It was precisely what I had in my head. The robotic didn’t transfer however the sky rotated as anticipated.
Remaining ideas
Total he did not do a foul job. What he did significantly properly was the unique picture, creating a way more lifelike illustration than some fashions obtain. The picture high quality was on par with some devoted AI imagers.
Issues arose with the motion of bigger objects. Whether or not it was a small motion or animating a component inside a big video, it labored fantastic. He had a tough time animating if the article was the dominant characteristic.
Haiper is a formidable addition to the AI video market, nevertheless it has a methods to go earlier than it reaches the longevity of Sora. For now, it additionally has a methods to go earlier than matching the movement consistency of Runway, StabilityAI, and Pika Labs fashions, nevertheless it’s very shut.
Having spoken to the crew behind Haiper, I’ve little question that they’ll broaden and resolve these issues in a short time. The primary downside is computing energy and coaching time, so with a current funding spherical of $13.8 million they’ll most likely improve to fulfill demand in a short while.
- MacBook Air M3 Introduced: Pre-Orders Now Open for ‘World’s Finest Client Laptop computer for AI’
- I examined Google Gemini vs OpenAI ChatGPT in a 9-round showdown – this is the winner
- OpenAI is constructing next-gen GPT-5 AI, and CEO claims it could possibly be super-intelligent
