I treated version numbers on AI video models as marketing. That held until I ran Alibaba’s HappyHorse 1.1, released June 22, 2026, and got back a 5.16-second clip at 1286 by 716 pixels. I was wrong about what a decimal point can fix. This piece covers what Alibaba says it changed, what showed on screen, and what the realism costs.
Alibaba Says 1.1 Targets Motion, Faces and Skin
Five areas upgraded, per Alibaba’s launch.
Alibaba released HappyHorse 1.1 on June 22, 2026, TechWeb reports. Compared with 1.0, the company upgraded motion, subject consistency, instruction following, visual texture and audio. It accepts up to nine character reference images at once, according to National Business Daily. CometAPI lists three modes: text to video, image to video and reference to video. The same report says 1.1 cuts the oily, over-sharpened look and keeps skin detail such as pores or smile lines.
The model runs on the HappyHorse site, Alibaba Cloud’s Bailian platform and Qwen Cloud, TechWeb says. Alibaba opened a global AI film contest the same day, with entries accepted until August 20. Entries needed at least 70 percent AI-generated footage, with at least 50 percent made in HappyHorse, TechWeb says. Gold-prize winners could earn up to $6,000 and 300,000 credits, per Sina.
CometAPI’s comparison pitches 1.1 for character consistency, reference-to-video and brand-controlled work. It pitches Seedance 2.0 for cinematic text-to-video.
My 5.16 Seconds Held One Crown Steady
Camera turns; the gold and gems stay put.
The file runs 5.16 seconds at 24 frames per second, about 124 frames, in H.264 at about 5 megabits per second. It carries an AAC audio track, consistent with the native audio that Tryonr advertises. Tryonr lists clips of 3 to 15 seconds at 720p or 1080p, so this file sits at the short end.
I pulled frames from across the clip. At about half a second, the man faces left in profile with a crowd behind him. By 4.6 seconds, he faces the lens with a hand raised while the crowd claps. That is a shift from side profile to front close-up in under five seconds.
The crown is what I watched. Its gold scrollwork and its pink and teal stones appear in both frames, though the angle changes. The face keeps the same beard and tight curls, and the skin shows creases instead of a smooth plastic finish. The street behind him is muted to near-grey while the crown stays saturated gold, which keeps the eye on it.
Behind him, the extras read as a real crowd. Men clap and smile, and one in a white cap has an arm raised. The setting rotates around him too: flagpoles and a stone building in the first frame give way to blurred trees and a grey tower in the last.
A Flipbook Needs Every Page to Agree
Consistency is 124 small agreements per clip.
A video model draws 24 new pictures for every second of footage in this file. Each frame lasts about 0.04 seconds, so one wrong crown flickers past before the eye names the problem. Subject consistency means each picture agrees with the one before it.
It works like a flipbook: if page 40 shows a different crown than page 39, the whole thing stops reading as one object. Reference images give the model something fixed to check against, which is why Alibaba’s nine-image limit matters for a character who must survive many shots. National Business Daily says the same input improves multi-shot understanding.
Real Skin Detail Makes Fakes Harder to Spot
The feature I liked also worries me.
Everything above has a price. A five-second clip is the unit of work, so a scene means many generations stitched together, and each one has to match the last. Keeping one face across them takes reference images, up to nine per generation, which is setup before any footage exists. The same detail that makes the skin look real also removes a visual tell that older, glossier output carried. Any clip I post now gets labeled as generated.
I now check whether the face and costume match from first frame to last before I judge anything else. I judged one 5.16-second clip, so I can’t say how often 1.1 holds a subject like this. A version number is worth one test clip.
