Open-Source · Apache 2.0
ERNIE Image is Baidu’s open-weight text-to-image model built on an 8B Diffusion Transformer. Engineered for precise text rendering, structured layouts, and complex multi-object prompts.
Deep Dive
ERNIE Image is an open-source text-to-image AI model developed by Baidu, built on an 8B-parameter Diffusion Transformer (DiT). It is designed to generate images with accurate in-image text, structured layouts, and complex multi-object compositions.
Compared to most open-weight models, ERNIE Image performs better on text-heavy and layout-sensitive tasks. It includes a built-in Prompt Enhancer that expands short inputs into richer, structured prompts, improving output quality without manual prompt engineering.
The model runs on a single consumer GPU with 24GB VRAM, making it suitable for local deployment. Released under Apache 2.0, it can be freely used, modified, and deployed commercially without API limits.
Core Capabilities
Six real capabilities that matter in production — not just model specs.
Produces sharp, readable text in posters, infographics, and UI-style images. Most diffusion models struggle with structured text, but ERNIE Image maintains clarity even in dense layouts. LongTextBench: 0.9733.
Builds consistent layouts across multi-panel designs, storyboards, and posters. Unlike typical models that focus only on visuals, ERNIE Image keeps layout logic intact. GENEval: 0.8856.
Accurately follows prompts with multiple objects, spatial relationships, and detailed instructions. Instead of collapsing complexity, it preserves structure across the entire scene.
Generates both photorealistic images and stylized visuals without switching modes. You can move from product shots to creative artwork in the same workflow.
Runs on a single 24GB GPU like RTX 3090 or 4090. No API, no cloud cost, and full control over your data and generation pipeline.
Expands short prompts into structured descriptions before generation. This reduces prompt engineering effort and improves output consistency.
Gallery
Real outputs that show where ERNIE Image performs best — especially in tasks most models struggle with.
A detailed pencil sketch of a pufferfish swimming inside a circular maze on the ocean floor, surrounded by seaweed, rocks, and bubbles.
Confident model wearing a bold blue and pink spiral-patterned suit with yellow shirt, heart-shaped yellow sunglasses, and pink earrings against a solid blue background.
Vibrant berry smoothie in a glass jar with dramatic splash of purple liquid, flying raspberries, blueberries and blackberries, cinematic lighting with a smartphone in the background.
A modern minimalist storefront shaped like a giant product can labeled 'BRAND PRODUCT', warm interior lighting, people walking outside on a city street at dusk.
Hand-painted watercolor sign on rustic paper in a forest, featuring a blue jay and flowers with text 'Native Wildlife: Please Observe from a Distance'.
Highly detailed technical blueprint of a gourmet smash burger with precise measurements, ingredient labels, and engineering specifications on a dark background.
Variants
Understand the key differences in quality, speed, and use cases — and choose the right version for your workflow.
The SFT model is the standard release — 50 denoising steps, full instruction fidelity, and the strongest benchmark scores. Use it for final renders where text accuracy and quality are non-negotiable.
GENEval 0.8856, LTBench 0.9733
ERNIE-Image-Turbo is a distilled variant trained with DMD. It cuts generation down to 8 steps — fast enough to preview 20+ compositions before committing to a final render.
Optimized for speed and exploration
| Capability | SFT (Main) | Turbo |
|---|---|---|
| Steps | 50 | 8 |
| Speed | Slower | ~6× faster |
| Best for | Final renders | Drafts, iteration |
| GENEval | 0.8856 | Lower |
| LongTextBench | 0.9733 | Lower |
| Available on | HuggingFace | HuggingFace |
FAQ
Quick answers to the most common questions about ERNIE Image.
Yes. ERNIE Image is free under the Apache 2.0 license.
You can download, use, modify, and deploy the model commercially without paying for API access or usage. There are no usage limits when running it locally.
ERNIE Image performs better at text rendering and structured layouts.
It outperforms most open-weight models in text-heavy tasks, while Midjourney focuses more on stylized visuals. ERNIE Image is better for posters, UI layouts, and readable text generation.
Yes. ERNIE Image supports commercial use under Apache 2.0.
You can use outputs for ads, products, and resale without additional licensing. Both the model and generated images are commercially usable.
ERNIE Image requires a 24GB GPU for the full model.
RTX 3090, RTX 4090, and A10G are commonly used. The Turbo version runs faster and may require less memory depending on your setup.
Yes. ERNIE Image works with ComfyUI out of the box.
You can load the safetensors checkpoint and use the official workflow template. It integrates with standard ComfyUI pipelines.
ERNIE Image supports English, Chinese, and Japanese prompts.
It can render bilingual text within a single image while maintaining readability. Performance is consistent across languages in benchmark tests.
Download model weights from Hugging Face, clone the official GitHub repository for setup and inference scripts, then run locally—or use the online demo in your browser when available.
Connect
Everything in one place — model weights, code, documentation, and the online demo.