ERNIE-Image Delivers Accurate Text-inclusive Image Generation

Baidu published ERNIE-Image, an open-source, single-stream Diffusion Transformer with 8B DiT parameters designed for high-fidelity text-in-image generation, layout fidelity, and bilingual prompts. The release includes a distilled 8-step ERNIE-Image-Turbo for draft-speed generation and a 50-step SFT model for final renders. Core benchmark numbers reported include GENEval 0.8856 and LongTextBench 0.9733, and the project is released under Apache-2.0 so weights and outputs are commercially usable.

ERNIE-Image is built on a single-stream DiT architecture and pairs the generator with a lightweight Prompt Enhancer that expands terse user inputs into structured descriptions. The team ships two runtime points: the distilled ERNIE-Image-Turbo (8 steps) for quick iteration and a full 50-step SFT model for quality. Supported resolutions include 1024x1024, and the codebase exposes generation, edit, composite, and upscale primitives so designers can centralize an asset pipeline.

Core capabilities: The model is explicitly trained to solve three failure modes common to diffusion-based image models: accurate long-form glyph rendering, instruction-faithful multi-object composition, and page-level layout reasoning. Independent and reproducible benchmark claims include LongTextBench 0.9733 for text rendering and GENEval 0.8856 for instruction fidelity, positioning ERNIE-Image ahead of several open competitors on text and layout-centric tasks.

Feature set:

• •Distilled 8-step ERNIE-Image-Turbo for rapid drafts and a 50-step SFT model for production frames
• •Built-in Prompt Enhancer to reduce manual prompt engineering and expand short prompts into structured instructions
• •Pipeline features: generate, edit, composite, upscale, and export in a single surface
• •Apache-2.0 licensing, enabling commercial use, fine-tuning, and self-hosting

ERNIE-Image addresses a persistent pain point for practitioners: the inability of many diffusion models to render legible, layout-sensitive text and strict compositions. Where Stable Diffusion and some proprietary models produce blurred or hallucinated glyphs, ERNIE-Image targets fidelity with dedicated training and evaluation on bilingual prompts. The Apache-2.0 license is strategically important for design teams and agencies that need clear commercial rights and the option to self-host or fine-tune. This release also reflects continued maturation of the Chinese foundation-model ecosystem, building on lineage work such as ERNIE-ViLG and signaling stronger competition in the open-weight image model space.

Practical implications for practitioners: If you operate creative pipelines, ERNIE-Image lowers the integration cost for generating posters, comics, storyboards, UI mockups, and ad frames that require accurate typography and layout constraints. Teams can self-host the weights, plug the Prompt Enhancer into existing prompt flows, and use the Turbo variant for iterative authoring followed by the SFT model for high-quality outputs.

Adoption by the open-source community, forks that add safety filters or domain-specific fine-tunes, and real-world asset pipelines that replace stock shoots. Also monitor compute costs for self-hosting, model safety and copyright moderation in downstream tooling, and how competitors respond on the text-in-image benchmark front.