POCKET-Image: Solving On-Device Korean Text Rendering

Existing diffusion models struggle with text rendering because they fail to learn Korean's complex jamo combination rules. POCKET-Image solved this by redefining the problem not as 'generation' but as 'preservation'.

Key Technology and Features:

• Glyph Fidelity Assurance: Instead of letting the model invent characters on its own, a pipeline was built to have the model reference accurate glyph shapes, maximizing multilingual text accuracy for Korean, Chinese, Japanese, and more.
• On-Device Optimization: Based on Z-Image (6B DiT), it applies 4-bit quantization (NF4, quanto) to dramatically reduce VRAM usage.
• VRAM Efficiency: Compared to the original (23.3GB), it can run at 8.6GB with NF4 quantization, and at around 4.5GB with CPU offload utilized, making it runnable on consumer-grade GPUs and Mac environments.
• Benefits of Quantization: Since text is preserved at the pipeline level, quantizing the model does not affect text accuracy, only having a minor impact on background quality.