CogVideoX is the open-source counterpart of the model behind QingYing. Five checkpoints span 720 by 480 to 1360 by 768, memory needs fall from 18GB to 3.6GB with INT8, and the 2B weights share the code's Apache 2.0 licence.

Opus Clip transcribes a long video, picks the passages worth publishing, reframes them for vertical and subtitles them, then schedules the batch across platforms. ClipAnything and ReframeAnything are its own models, and every clip gets a virality score.

The desktop build of CapCut puts one-tap AI next to manual tools: smart cutout, grading, quality repair, frame interpolation and beauty sit beside four kinds of mask, keyframes and multicam. New hardware support covers 4:2:2 encoding and RTX 50 series acceleration.

One portrait plus a script is the whole input: no scan, no capture rig, nobody on set. The same face then holds across more than 120 languages, and a second product turns it into something that answers questions live.

The expensive part of an induction or compliance video is rarely the shoot. It is finding the presenter again, shipping six language versions and getting past legal, which is exactly the loop a browser-based studio with 240 avatars is aimed at.

Fifteen seconds of webcam footage is enough to build a presenter that survives a change of set, a change of clothes and a change of language. What makes that work is not a better face but an identity signal lifted across frames and kept apart from appearance.

Two experts of about 14B each add up to 27B, yet only one fires per step, so the cost stays near a single-expert model. A separate 5B branch keeps 720P at 24fps inside a single consumer card, and compression in the VAE is what buys the room.

Effects, ambience and spoken lines come out of one pass instead of being laid on afterwards, and the clip lands in 1080p or 4K with the frame, the movement and even objects inside it editable. A studio layer wraps the model in tools and a daily credit allowance.

One generation now yields a quarter-minute clip at full HD with several shots and its own dialogue, effects and ambience, and the frame shape is fixed before the render instead of cropped after. A separate line pushes the same engine toward worlds that keep answering input.

Dialogue, narration, effects and music leave in the same export as the picture, so a scene no longer needs a silent render followed by dubbing. A second line handles the live side, where an avatar talks back and scenery swaps without stopping the stream.

The generator answers prompts, stills and audio in the same pass, keeps stereo in the render rather than adding it later, and treats 2K as the starting point. Its cheapest claim is a per-second rate the vendor puts below a third of what mainstream models charge.

Sound is settled in the same pass as the image, so a clip does not need a silent render followed by dubbing. Native 4K sits at the top of the price ladder, and the developer side charges by the second.

A single clip can carry up to 16 keyframes, existing footage can be rewritten for 20 seconds at a time, and 16-bit colour arrives straight from the model. That is the argument: AI output walking into a normal finishing pipeline.

Stills, effects, trend imitations, an agent and an interface all live on one site, and the free plan hands out 80 credits a month. A five second frame at 1080p costs 40 of them, so the free allowance buys two.

Generation, image work and finishing share an account and a workspace, so a clip does not have to be exported to be repaired. The free plan hands over 125 credits once, and paid plans start at USD 12 a month billed yearly.