Styled Text Image Generation with Eruku on AMD

2026-04-24 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, long

Summary

Eruku is a styled text image generation model, presented at WACV 2026, that addresses the challenge of generating readable and controllable text images while faithfully matching a target visual style. It utilizes an autoregressive encoder-decoder Transformer architecture, contrasting with diffusion models, and employs an explicit end-of-generation token for improved reliability. The model was trained on a 20 million sample synthetic dataset and fine-tuned on a smaller dataset of longer samples, demonstrating competitive performance against predecessors like Emuru and diffusion-based methods like DiffusionPen. Eruku was developed and trained on the LUMI supercomputer using AMD Instinct™ MI250X GPUs and ROCm™, with its code compatible with both CUDA and ROCm™ for flexible development.

Key takeaway

For AI Engineers and ML Scientists working on text image generation, Eruku offers a robust autoregressive solution for high-fidelity styled text. You should consider deploying Eruku on AMD Instinct™ MI250X GPUs with ROCm™ for scalable training and inference, especially for tasks requiring precise control over text content and visual style, such as synthetic document creation or handwriting synthesis.

Key insights

Eruku is an autoregressive Transformer for styled text image generation, outperforming diffusion models in text control and style fidelity.

Principles

• Autoregressive models excel in variable-length text image generation.
• Explicit end-of-generation tokens enhance content adherence.
• Synthetic datasets can effectively train robust text generation models.

Method

Eruku tokenizes text with ByT5 and processes style images via Emuru's VAE encoder. An encoder-decoder Transformer then generates image tokens autoregressively, stopping upon an end-of-generation token, which are then decoded by the VAE.

In practice

• Use Eruku for synthetic handwriting generation.
• Apply Eruku for graphic design text styling.
• Integrate Eruku for OCR training data creation.

Topics

• Eruku
• Styled Text Image Generation
• AMD Instinct MI250X
• ROCm
• LUMI Supercomputer

Code references

• Blowing-Up-Groundhogs/Eruku
• Lumi-supercomputer/LUMI-AI-Guide
• Blowing-Up-Groundhogs/Eruku
• huggingface/accelerate
• spaces/carminezacc

Best for: Machine Learning Engineer, AI Engineer, AI Scientist

Related on AIssential

• Accelerating ComfyUI Workflows on AMD Instinct™ MI355X GPUs with ROCm — AMD Instinct MI355X GPUs significantly accelerate ComfyUI generative AI workflows over NVIDIA B200.
• OpenAI Beefs Up ChatGPT’s Image Generation Model — OpenAI's ChatGPT Images 2.0 enhances image generation with multi-image and multilingual text capabilities.
• Efficient and Portable 3D Explorable World Generation on AMD GPUs — Optimizing 3D world generation on AMD GPUs requires kernel portability and pipeline efficiency.
• AI, Design, and the Power of Open Models — Image generation models must prioritize user control, accurate text, and customizability for professional design workflows.
• Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models — Diffusion Language Models (DLMs) enable parallel, iterative text generation and revision, significantly boosting speed and accuracy over autoregressive models.
• KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR — Current MLLMs struggle with OCR and language identification for low-resource Kazakh scripts, especially Arabic, performing worse than traditional OCR.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.