← Back to blog
AiAbout 7 min read

Chinese text in video always turns to mush. Dreamina handed the typesetting to code

Published Nov 15, 2026
Chinese text in video always turns to mush. Dreamina handed the typesetting to code

Anyone who makes explainer videos knows the same pain: however good the footage looks, the text inside it is still blurry. A three-character title drifts into six, subscripts in a formula slide out of place, and a data label becomes a different number by the second mark. To keep subtitles legible, everyone ends up doing the same thing: use the generated footage as a background, open CapCut, and stack real text on top, layer by layer.

Dreamina's Motion Studio, launched on November 10, takes a different route. It does not generate continuous frames. It has the model write code, and the code arranges the layout, entrance order and motion rhythm of text, images and shapes. The official summary: describe the idea in words, compile the visuals through code.

That sounds like marketing copy, but the technical meaning is clear: the text is rendered by a layout engine, not painted by a diffusion model. Painted characters blur. Typeset characters do not.

A 60-second video: 8.5 minutes, 82 credits

Start with the numbers, because they decide whether this fits into a daily workflow.

Motion Studio is reached by signing into Dreamina in a desktop browser and using the entry on the left of the home page. You can set aspect ratio and duration, the estimated credit cost appears before you generate, and the finished file downloads directly. Dreamina has not published separate pricing for the feature; it runs on credits.

Jiqizhixin's measurements are specific: a 60-second explainer took about eight and a half minutes and 82 credits; a 3-minute video took about 13 minutes. The rhythm is closer to "go make tea, come back and revise the second cut". One click and out does not happen here.

The wait is reasonable, because more happens than generating a clip. The pipeline first parses the intent and plans, then generates voiceover and visuals separately, composites and renders, and finally runs a check on the finished video. Four stages inside one click; eight and a half minutes is not slow.

Why frame-by-frame generation keeps losing on Chinese

To see the value of this route, accept one fact first: when a video model draws Chinese, it is drawing shapes, not writing characters.

Diffusion models produce text through pixel distributions. Glyph forms seen in training data can be imitated; unseen combinations get guessed. English words are short, have few letters and many common combinations, so the guess often passes. A Chinese character is a complex structure on its own, dense in strokes and components, and the model has no concept of "character", only of "this patch should be ink". The result reads as text from a distance and as nonsense up close.

That explains why so many AI video thumbnails look premium until you zoom in on the gibberish. It also explains why teams in China making AI short dramas and AI talking-head videos still end up retyping subtitles in an editor. This is not a taste problem. The generation method simply does not match the object.

Motion Studio replaces that layer. Text goes to code-based typesetting, where typeface, size, weight, leading and line breaks are set by layout rules, and the model only decides when something appears, where it enters from, and how long it stays. Once the division of labour changes, blurry characters stop being a problem at the root.

What code layout solves, and what it doesn't

Clarity and control. Titles, formulas and data labels, anything that needs a definite glyph, now has a definite rendering path. This matters especially for Chinese content, where information density is high, many characters share one screen and there is little room for error.

Control matters more. Frame-by-frame generation is a gacha pull: you want the title to slide in from the left at second three, the model may fade it up from the bottom, and three re-rolls cost real money. Code arrangement is reproducible. Change one parameter and that parameter moves; nothing else follows. Iteration works the same way: fix the broken part on its own instead of redoing the whole video.

What it does not solve is whether the content is right. Jiqizhixin's test is blunt about this: in a demo about the evolution of Chinese characters, the first cut contained wrong glyphs, corrected only after reference material was supplied. Code guarantees the characters are laid out clearly, not that they are written correctly. For material with specialist knowledge, rare characters or terminology, creators still have to watch frame by frame.

One judgement of my own: many people will miss this boundary. The better a tool works, the easier it is to assume the output is correct. Motion Studio solves "the footage doesn't look AI-made". The remaining question, "does the content look human-made", is beyond it.

Its real rival isn't Seedance

By product positioning, Motion Studio and the Seedance line are not substitutes. Seedance aims at something close to live action: light, material, natural motion. Motion Studio works in graphics and layout, emphasising style, rhythm and the marriage of sound and image. One shoots footage; the other sets a page.

The real rival is CapCut's template library, and workflows that pair tools like Canva and Figma with manual layout. They serve the same need: I want a short clip with rhythm, legible text and fitting music. Their strength is certainty; their weakness is that you build it by hand every time.

Motion Studio enters at "one sentence in, structure builds itself". Give it a noun and it assembles the explanatory structure, the pacing and the voiceover. The template side covers character animation, retro pixel, geometric motion and creative collage, with visual languages spanning geometric abstraction, ASCII, pop and Y2K; pick one and example prompts fill in automatically, then change theme, copy and palette, or upload your own assets.

The combination suits one kind of content particularly well: explainers. The "vibe knowledge" videos circulating on Douyin and Bilibili, which turn abstract ideas like entropy or recursive self-improvement into narrated, scored animation, are exactly the target. High information density, clear visual requirements, no need for live-action texture: the case where frame-by-frame generation is least economical and code layout most.

Studio-ification: packaging the workflow into a product

Before Motion Studio, Dreamina shipped a short-drama studio on October 26, and says a marketing studio for e-commerce and brand work is coming soon.

Read together, the logic is plain: sell outcomes by scenario instead of selling capability by model. Users should not need to know the difference between Seedance and a motion engine, only that they want a short drama, an explainer, or a product shot. The product absorbs the rest.

It is the common choice for Chinese video tools this year. As model capability converges, whoever has the shorter path, fewer jumps, fewer exports and imports, gets kept. Motion Studio compresses four or five steps, model writes code, browser runs it, screen-record and export, add voiceover and music, into one click.

Where I'm still not convinced

First, eight and a half minutes for a 60-second video is still slow for volume work. Teams shipping dozens a day either run several jobs in parallel or save the tool for hero pieces.

Second, credit transparency is thin. A pre-generation estimate is good, but what 82 credits cost depends on which plan you bought, which makes per-video cost hard to state when you are planning a schedule.

Third, and most practical: template-driven motion videos look alike. Once enough people use the geometric, ASCII and Y2K sets, viewers recognise the tool in three seconds. The platform gives a starting point; differentiation is still yours to add.

If you want to try it, start with a 60-second explainer on a subject you know well and the model does not, say the internal process of some industry. You will see two things at once: whether the typesetting is genuinely clear, and how badly the content can be wrong.

Related articles