ERNIE Image Turbo is a distilled text-to-image generation model developed by Baidu's ERNIE-Image team. Released as an open-source model under the Apache 2.0 license, it is optimized for high-speed inference while maintaining the visual quality and instruction-following capabilities of the base ERNIE-Image series. It is designed to generate high-resolution images in approximately eight inference steps, significantly reducing latency compared to standard diffusion models that typically require 50 or more steps.
The model is built on a single-stream Diffusion Transformer (DiT) architecture featuring 8 billion parameters. Its efficiency is achieved through a combination of Diffusion Model Distillation (DMD) and reinforcement learning (RL) techniques, which condense the generation pipeline without a substantial loss in detail or aesthetic quality. The system also utilizes a secondary 3B parameter prompt enhancer that automatically expands brief user descriptions into richer, more structured prompts to better leverage the model's creative capabilities.
A core strength of ERNIE Image Turbo is its proficiency in text rendering and structured visual generation. It supports native prompt understanding for Chinese, English, and Japanese, and has shown strong performance in rendering dense, layout-sensitive text within generated images. These capabilities make it particularly suitable for specialized tasks such as creating commercial posters, multi-panel comic layouts, and infographics where accurate typography and precise multi-object relationships are required.
For optimal results, the model is typically paired with the official prompt enhancer to handle brief inputs. It performs effectively on consumer-grade hardware, requiring between 12GB and 24GB of VRAM depending on the precision and optimization used. The model's training focuses on balancing visual appeal with controllability, ensuring that accurate content representation matches the aesthetic output.