Text and fonts in any language

Add every word of on screen text in post, with a font file you have checked against your real copy. A font that "has the letters" is not enough: it must also shape them, your render must use that exact file, and the layout must follow the rules of the script. This page works for any language and any tool. If you make video with code, also read Fonts in code video.

Coverage is not supportCopy Anchor Link

A font can contain every character of a script and still draw the language wrong. Google's shaperglot project gives the classic case: a font can cover the whole Arabic block (U+0600 to U+06FF) and still fail Arabic, because Arabic letters need joining forms (initial, medial, final). Devanagari needs conjuncts and reordering. Turkish needs the dotted and dotless I to change case correctly.

So a real check has two parts:

• Coverage: the font maps every character in your real copy.
• Shaping: the font joins letters, attaches marks, reorders and builds conjuncts the way the language needs.

One language can need several subsetsCopy Anchor Link

Google Fonts and Fontsource cut each family into subsets, and a language often needs more than one. On 15 Oct 2026, Roboto at weight 400 came back as nine subsets: cyrillic-ext, cyrillic, greek-ext, greek, math, symbols, vietnamese, latin-ext and latin.

Load only latin for Vietnamese and some letters fall back to another font, so one word is drawn in two fonts. The Google Fonts text= option returns a file cut to just the characters you pass, which Google says can make it up to 90% smaller. That suits a title card whose text never changes.

How to check a font for your languageCopy Anchor Link

# coverage.py FONT COPY.txt : list characters in the copy that the font cannot map
import sys, unicodedata
from fontTools.ttLib import TTFont
font = TTFont(sys.argv[1], fontNumber=0)
text = unicodedata.normalize("NFC", open(sys.argv[2], encoding="utf-8").read())
cmap = font.getBestCmap()
missing = sorted({c for c in text if not c.isspace() and ord(c) not in cmap})
print("missing:", " ".join(f"{c} U+{ord(c):04X}" for c in missing) or "none")
sys.exit(1 if missing else 0)

pip install fonttools brotli hyperglot
python coverage.py BrandDisplay.ttf copy_vi.txt
hyperglot BrandDisplay.ttf
cargo install --git https://github.com/googlefonts/shaperglot
shaperglot check BrandDisplay.ttf vi ar hi

A real example from the research for this page: macOS Futura has đ and ê but not ầ or ư. In the word "đầu", the first letter would come from Futura and the second from a fallback font. That is the ransom note look inside a single word.

Noto as the safe fallbackCopy Anchor Link

Google's Noto covers more than 150 writing systems and over 1,000 languages, with one family per script: Noto Sans Thai, Noto Sans Arabic, Noto Naskh Arabic, Noto Sans Devanagari, Noto Sans Hebrew and so on. When your brand font lacks a script, the matching Noto family is the default partner.

For Chinese, Japanese and Korean, pick the region variant: JP (Japanese), KR (Korean), SC (Simplified Chinese), TC (Traditional Chinese) or HK (Hong Kong). The three languages share Han characters, but readers expect small differences in the shapes. A Japanese video set in Noto Sans SC is correct Unicode and still looks wrong to Japanese readers. Set the language tag too (for example lang="ja"), because some browsers pick glyph shapes by language.

Fallback without the ransom note lookCopy Anchor Link

When a font in your list lacks a character, the renderer tries the next font, one character at a time. If no font you supplied has it, it uses whatever is installed on the machine. That gives two failures:

• Tofu: empty boxes. No font in the list or on the machine has the glyph.
• Ransom note: the glyphs exist, but come from different fonts letter by letter or word by word, with different sizes, weights and baselines.

The usual causes of the ransom note are loading only the latin subset, a display font with partial coverage, and decomposed text: a base letter followed by a separate combining mark, which some fonts cannot draw as one letter. Normalize your copy to NFC (precomposed letters) before you render, especially for Vietnamese, unless you know the font handles decomposed marks.

Pick a pattern per languageCopy Anchor Link

To match a partner font, pick one with similar proportions, then adjust its size until both sit at the same visual height on one line. In CSS this is size-adjust; in an editor, nudge the size by eye. When the brand font cannot cover a script at all, set that language's headlines entirely in the partner font and accept that some brand cues are lost. HyperFrames' own style presets do this for Chinese and note that the uppercase look of the Latin design disappears.

Starting fonts per languageCopy Anchor Link

Script rulesCopy Anchor Link

Each script has a few rules that break motion design habits. Check the rows for every language in your video.

Right to leftCopy Anchor Link

• Set the base direction at the top of the layout, not letter by letter. In HTML that is dir="rtl" on the root element. W3C advises against setting the base direction with CSS.
• Align with start and end instead of left and right, so one template works in both directions.
• Mirror anything that shows order or progress: progress bars, wipes, slide ins, typewriter reveals. Do not mirror logos, universal signs, photos, clocks or the digits inside a number (Apple Human Interface Guidelines). Reverse a sequence of numbers only when it shows progress.
• Hebrew uses Western digits. Arabic uses Western or Eastern Arabic digits depending on the region.
• Apple suggests setting Arabic and Hebrew slightly larger to balance uppercase Latin on the same screen.
• Mixed lines (Arabic or Hebrew with Latin words and numbers) are where bidi bugs show up. WebVTT takes each line's direction from its first strong character. Add U+200F (right to left mark) or U+200E (left to right mark) at the start when that guess is wrong.

Stacked marks and line heightCopy Anchor Link

W3C notes that non Latin text is often much taller than Latin: its Thai example needs about 150% of the vertical space of the English one. Vietnamese stacks a tone mark over a vowel mark (ấ, ở, ữ) and Thai stacks vowel and tone marks above consonants. Display line heights of 0.9 to 1.0 clip those marks or make lines touch.

The starting values in the table above are our synthesis of W3C and type design sources, not a standard. Test with the tallest string in your copy, for Vietnamese an uppercase line with stacked marks such as "ẤN TƯỢNG". Masks and wipes that reveal text clip marks too, so give them extra room at the top.

Text expansion when translatingCopy Anchor Link

W3C quotes IBM's guidance for translating from English into European languages:

The same article measured the Flickr label "views" at 0.8 times the English width in Korean, 1.2 in Chinese, 2.8 in German and 3 in Italian. One word titles and buttons grow the most, so size boxes for the longest language, not for English.

UppercaseCopy Anchor Link

All caps follows language rules only when the language is set, and support varies between browsers.

Dutch ij becomes IJ. If your tool has an all caps switch but no language setting, type the capitals yourself and check them.

Captions and burn inCopy Anchor Link

Pick a formatCopy Anchor Link

Burn in with FFmpeg and libassCopy Anchor Link

FFmpeg's subtitles and ass filters use libass, which handles shaping, right to left text and line breaking when it is built with the right libraries.

# Check that this ffmpeg has the subtitle filters
ffmpeg -hide_banner -filters | grep -E ' (subtitles|ass|drawtext) '

# Burn an SRT in any language with shipped fonts, complex shaping and Unicode wrapping
ffmpeg -i in.mp4 -vf "subtitles=subs.th.srt:fontsdir=./fonts:shaping=complex:wrap_unicode=1:force_style='Fontname=Noto Sans Thai,Fontsize=22,Outline=2,MarginV=60'" -c:a copy out.th.mp4

# Burn a styled ASS file
ffmpeg -i in.mp4 -vf "ass=subs.ar.ass:fontsdir=./fonts:shaping=complex" -c:a copy out.ar.mp4

Right to left captions in ASSCopy Anchor Link

libass uses left to right as the base direction by default, for compatibility with the old VSFilter renderer. For Arabic, Hebrew, Persian and Urdu, set the style's Encoding field to -1, which tells libass to detect the direction. Add a Language: line under [Script Info]; libass reads the first two letters as a hint for language specific glyph forms, such as Serbian Cyrillic. Keep Spacing at 0 for Arabic, and Italic at 0 because Netflix's Arabic guide says not to use italics in Arabic.

[Script Info]
ScriptType: v4.00+
PlayResX: 1920
PlayResY: 1080
Language: ar

[V4+ Styles]
Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
Style: Default,Noto Sans Arabic,58,&H00FFFFFF,&H00FFFFFF,&H00000000,&H64000000,0,0,0,0,100,100,0,0,1,3,0,2,80,80,70,-1

Line length per scriptCopy Anchor Link

Netflix's timed text rules are a good reference, even though social captions are usually larger (see Finishing for sizes and safe zones). General rules: 2 lines at most, each caption on screen from 5/6 of a second to 7 seconds, and lines broken after punctuation and before conjunctions or prepositions.

Captions in editorsCopy Anchor Link

• Premiere Pro: until late 2021, Hebrew and Arabic SRT files imported reversed unless you switched the caption text engine to "South Asian and Middle Eastern". Since version 22.1 (December 2021) a Universal Text Engine handles this, according to an Adobe community thread. Edge cases remain: a 2024 report says Hebrew punctuation does not turn right to left in MOGRTs.
• CapCut: we found no official documentation on Arabic, Indic or Thai shaping, right to left text or custom fonts for non Latin scripts. Treat it as unverified for those scripts and check every export.
• Safe fallback for any editor: render the text in a browser engine, which shapes, wraps and handles direction correctly, as a transparent overlay, then drop it on your timeline.

# HyperFrames: transparent WebM, or ProRes 4444 MOV for editing (page background must be transparent)
npx hyperframes render --format webm
npx hyperframes render --format mov

# Remotion: ProRes 4444 with alpha
npx remotion render --image-format=png --pixel-format=yuva444p10le --codec=prores --prores-profile=4444

Text inside AI generated video and imagesCopy Anchor Link

No video or image model publishes how accurate its text is per script. Research on text rendering explains why: many models do not see individual letters, and even systems built for text still struggle with many languages. Latin gets the most attention, then Chinese, Japanese and Korean. Arabic, Indic and Thai have the fewest vendor claims and the most creator complaints.

What the vendors say (checked 15 Oct 2026):

Seedance subtitle syntaxCopy Anchor Link

Seedance 2.5 has a sound and text syntax: () for music, <> for sound effects, {} for dialogue and 【】 for subtitles. For dialogue that is not in Chinese, ByteDance recommends naming the language before the line. To avoid unwanted subtitles, its guide suggests:

• Write "No subtitles." in the prompt. One official example uses "No subtitles, no text overlays".
• Do not repeat dialogue words elsewhere in the prompt, and do not attach tone instructions to single words.
• Write lines as "Character's line (emotion): content".
• Use reference videos that contain no subtitles, since they can override your instruction.
• Keep text off storyboards you upload.

Subtitles made with 【】 are fine for drafts, in languages on Seedance's list. They are burned into the pixels in a font you do not control, so you cannot fix a typo, restyle them or translate them later. See Seedance for the rest of the model.

The rule: add text in postCopy Anchor Link

Night street, light rain, a small noodle shop with warm light inside.
Above the door, one shop sign that reads exactly "Phở Hà Nội" in white letters,
matching the attached reference image. No other text anywhere in the frame.
No subtitles, no text overlays, no captions.

LicensingCopy Anchor Link

• Subsetting a font or converting its format counts as a change under the OFL. That is allowed, but if the font declares a Reserved Font Name, rename the family in your subset before you share it. Wrapping the font in WOFF without changing its data is the exception.
• Videos you rendered with Adobe Fonts keep working after you cancel the subscription.
• Some tools copy installed fonts into the project automatically. HyperFrames local renders do this, which can put a commercial system font into a file you share. Check the license before that happens.

ChecklistCopy Anchor Link

• Write down each language's tag, script, direction, whether it uses spaces, has case, is cursive and stacks marks.
• Collect the final copy for every language, normalized to NFC, including the longest strings.
• Run a coverage check on the real copy and a shaping check; pick a partner font per script and a region font for Han.
• Use real weights only, and a license that allows your video, your render machine and sharing the files.
• Set the language on all text, set the direction for right to left, and mirror directional motion only.
• Tracking at 0 for Arabic and Indic scripts, no forced uppercase on caseless scripts, uppercase with the language set.
• Raise line height for stacked marks and pad reveal masks.
• Let the renderer wrap lines; never split Thai, Chinese or Japanese on spaces.
• Animate complex scripts by word or line.
• Burn captions with a libass build of FFmpeg: fontsdir, shaping=complex, wrap_unicode=1, and Encoding -1 for right to left.
• Generate AI footage without text and add all text in post.
• Have a native reader check every frame that shows text.

What creators report on XCopy Anchor Link

Searches run on 15 Oct 2026. Labels: organic means a creator sharing their own result; self promo means a creator promoting their own free tool; vendor means the maker; sponsored means a paid post.

• Seedance 2.5 writing Japanese on a blackboard still makes character mistakes, and long sentences and small text break easily. Organic, @tanabe_fragm, 12 Aug 2026.
• A creator had Seedance 2.5 generate Japanese dialogue with English subtitles and says it even sized the text on its own. Organic, @CuriousOne_01, 21 Aug 2026.
• Chinese prompt writers routinely end Seedance prompts with negatives for subtitles (字幕) and garbled text (乱码), for example @liyue_ai and @xianshi666123. Organic.
• When a sign must be in the scene, one prompt asks for exactly one green Chinese exit sign (安全出口) and forbids English or garbled text. Organic, @johnAGI168, 26 Aug 2026.
• A Seedance concept ad describes the brand end card, then says the brand text is overlaid in post to keep it clear and accurate. Organic, @hxhxhx0916, 4 Oct 2026.
• A commentator says Alibaba's docs for a new document to video model tell users to composite legal copy and subtitles afterwards. Secondhand, we did not verify that doc. @lordofblocks, 2 Sep 2026.
• Veo 3 made errors in Japanese text. Organic, @okloorcl, 25 May 2025.
• Nano Banana Pro drew Arabic with diacritics well, but the final image still took a few more attempts and half an hour in Photoshop. Organic, @Ibraheem_Abbas, 21 Nov 2025.
• With the same prompts in GPT Image 2.5 and Nano Banana 2.1, Nano Banana added text to a poster that was not in the prompt. Organic, @olezhavich, 9 Oct 2026.
• A post praising GPT Image 2.0 on Higgsfield for accurate Japanese text ends with "PR". Sponsored, @KanaWorks_AI, 22 Apr 2026. Many resellers repeat "clearer text" claims for Nano Banana; treat them as marketing.
• Captions in Arabic and Moroccan Darija: letters do not always display well, right to left text causes problems, and mixing Arabic with French and numbers is hard. Organic, @ElmaAbdeljalil, 13 Oct 2026.
• Vietnamese viewers notice broken fonts: one mocked a promo video for its "lỗi font" (font errors). Organic, @NcTn_1611, 5 Sep 2024.
• CaptionSpec checks SRT files against Netflix rules (reading speed, line length, timing) in the browser. Self promo, @CreateLex, 12 Oct 2026.

The pattern: creators trust code and editors for text and treat model text as decoration. Complaints cluster on Japanese and Arabic in generated frames, and on right to left text mixed with other scripts in captions.

SourcesCopy Anchor Link

• shaperglot and Hyperglot
• Google Fonts CSS2 API, Google Fonts FAQ and Fontsource subsets
• Noto fonts, noto-cjk and W3C, why use the language attribute
• CSS Fonts Level 4 and CSS Text Level 3
• MDN, size-adjust, MDN, font-synthesis, MDN, letter-spacing and MDN, text-transform
• W3C Indic layout requirements, W3C, text size in translation and Vietnamese Typography, diacritical details
• W3C, structural markup and right to left text, Apple, right to left and WebVTT
• Chrome, CSS i18n features
• FFmpeg filters, Homebrew ffmpeg, Homebrew ffmpeg-full and libass shaper source
• Netflix, characters per line, Netflix general timed text guide and Netflix Arabic guide
• Adobe community, reversed RTL captions in Premiere and CapCut help, text not showing
• HyperFrames rendering guide and Remotion transparent videos
• Seedance 2.5 docs and Seedance 2.5 prompt guide
• Vertex AI video prompt guide, Wan2.1, OpenAI image generation and Gemini image generation
• Character-aware models improve visual text rendering and Glyph-ByT5-v2
• OFL FAQ, OFL text, Adobe Fonts licensing and Apple fonts