DEV Community

AI OpenFree
AI OpenFree

Posted on

Why Text in AI-Generated Images Is Always Broken — and the On-Device Model That Fixes It

#ai #imagegeneration #opensource #korean

Why Text in AI-Generated Images Is Always Broken — and the On-Device Model That Fixes It

If you've ever tried to make a poster with an AI image model, you know the feeling. The background looks gorgeous. The composition is clean. And then you look at the text — and it's gibberish. You asked for "안녕하세요" and got "안ㅐ기". You asked for a brand name and the letters melt into each other.

This isn't a skill problem. It's a structural one. And POCKET-Image, from VIDRAFT, is built to fix it.

Why global models garble non-Latin text

Most world-class image models are trained with Latin script at the center of gravity. When they meet a writing system with complex strokes or combinatorial rules — Korean Hangul, Chinese, Arabic, Thai — they don't really write the characters. They paint something that looks like the characters. The result is always almost-right, which in practice means wrong: you end up reopening Photoshop to lay the text down by hand.

Korean is a hard case on purpose. Hangul composes 자모 (jamo) into 11,172 syllable blocks, and the shape of each component shifts depending on where it sits in the block. A model that learned to "draw" text rather than to compose it simply can't reproduce that reliably.

What POCKET-Image does

POCKET-Image is an on-device "text-in-image" system. You describe a scene, give it the words you want, and it writes those words without breaking a single character.

  • Supports 한국어 · 中文 · 日本語 · العربية · ไทย · Latin and more
  • Handles right-to-left (Arabic) and complex shaping (Thai, Indic scripts)
  • Text is optional — leave it empty and you get a pure image

So instead of "generate a nice background, then bolt text on top," you get an image where the text is correct from the first render. That matters most exactly where text is the product: card news, posters, signage mockups, thumbnails, and multilingual marketing.

No GPU. No cloud.

The second thing that makes POCKET-Image different is that it runs on-device.

  • No GPU, no NPU — it runs on a plain CPU and RAM
  • From 4.5 GB — Windows, macOS, Linux, and Apple Silicon
  • Fully local — nothing goes to the cloud, no account required

And here's the number that matters most. The same model needs about 23 GB of memory at full precision. POCKET-Image runs it in as little as 4.5 GB — roughly 5.2× smaller — with no visible loss in quality. That's the whole POCKET idea: shrink a big model until it fits on the hardware people actually own. 4.5 GB is a 6 GB consumer GPU (a GTX 1660 / RTX 2060) or a laptop — not a data center.

Design source material is often confidential before launch. A new product poster, an unannounced brand name, an internal campaign — finishing those inside your own laptop, without a round trip to someone else's server, isn't just about speed or cost. It's about security and data sovereignty.

Built on open foundations

POCKET-Image is built on Z-Image (Apache-2.0), a commercially-friendly open foundation, combined with VIDRAFT's on-device POCKET-Core engine. The model itself is released under Apache-2.0 — download it, run it, modify it, redistribute it. No strings.

Try it right now

If installing sounds like a hassle, there's a browser studio. Type your text and prompt, and it generates the image on the spot.

Part of the POCKET family

POCKET-Image is the image branch of VIDRAFT's on-device family, POCKET — built on one principle: big models, small hardware, no cloud. Alongside the chat lineup that runs a 35B language model on a GPU-less PC and on phones (POCKET-35B / 26B / KR / EN), POCKET now brings one of the oldest unsolved problems in image generation — correct text, in any language — on-device too.

No cloud, no GPU, no account — every glyph, exactly right. That's POCKET-Image.

Built by VIDRAFT — a Pre-AGI AI company working on foundation models, on-device inference, and applied AI tools.

Top comments (0)