---
license: apache-2.0
---

# baidu/ERNIE-Image Model Cards

## Model Details

### Model Description

**ERNIE-Image** is a text-to-image generation model developed by the ERNIE team at Baidu.

In terms of image quality, ERNIE-Image is on par with current state-of-the-art models. It demonstrates significant advantages in handling complex instructions, particularly in tasks that require **accurate text rendering** and **knowledge-intensive generation**.

### Key Features

* **Precise text rendering**: Especially strong in dense or complex text scenarios
* **Excellent instruction following**: Accurately interprets and executes complex prompts
* **High-quality portraits and stylized images**: Strong performance in both realism and artistic styles

### Model Architecture

ERNIE-Image consists of the following components:

* An 8B-parameter Diffusion Transformer (DiT)
* A 3B text encoder from Ministral
* A VAE based on flux2.dev
* A prompt enhancer fine-tuned using Ministral 3B

### Deployment

Thanks to its relatively compact model size, ERNIE-Image can be deployed on **consumer-grade GPUs (e.g., 24GB VRAM)**, making high-quality image generation more accessible and practical.

## Evaluation

### Benchmark


### Showcase
<style>
.masonry {
  column-count: 2;        /* 两列 */
  column-gap: 12px;
}

.card {
  break-inside: avoid;    /* 防止卡片被切断 */
  margin-bottom: 12px;
}

.card img {
  width: 100%;            /* 宽度统一 */
  height: auto;           /* 高度自适应（关键） */
  display: block;
  border-radius: 8px;
}
</style>

<section id="shape-3160722256177163633" class="tab-panel active">
  <div class="masonry">
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/ukAjGbYZwG4jPRRs9UgJ0.jpeg">
    </article>
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/nD1aI60oXzAuGnhBA0P8T.jpeg">
    </article>
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/YveUuXuSIiBWDs-U_mYqI.jpeg">
    </article>
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/_0RGqPUeIB0r5SigsPgjz.jpeg">
    </article>
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/RLQonbN1cRg4cHyx3GnCS.jpeg">
    </article>
    <article class="card">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/6358acfe0e4fef21982a929b/C8NxWvTeC0-tqZuBmVPMT.jpeg">
    </article>

  </div>
</section>

## Uses

### Installation & Download
Install the latest version of diffusers:
```
pip install git+https://github.com/huggingface/diffusers
```
Download the model:
```
pip install -U huggingface_hub
HF_XET_HIGH_PERFORMANCE=1 hf download baidu/ERNIE-Image
```
### Recommended Parameters
- Resolution: 
    - 1024x1024
    - 848x1264
    - 1264x848
    - 768x1376
    - 896x1200
    - 1376x768
    - 1200x896
- Guidance scale: 4.0
- Inference steps: 50
  
### Usage Example
```
import os
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
import random
import numpy as np
import torch
from diffusers import ErnieImagePipeline

seed = 42
print(f"seed: {seed}")
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.use_deterministic_algorithms(True)
torch.backends.cudnn.benchmark = False

# 加载 pipeline
pipe = ErnieImagePipeline.from_pretrained(
    "baidu/ERNIE-Image",
    torch_dtype=torch.bfloat16,
)
pipe = pipe.to("cuda")
pipe.transformer.eval()
pipe.vae.eval()
pipe.text_encoder.eval()
pipe.pe.eval()
# 如果是消费级显卡,例如 Nvidia 3090
# pipe.enable_model_cpu_offload()

# 设置随机种子
generator = torch.Generator(device="cuda").manual_seed(seed)
# 生成图片
output = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=5.0,
    generator=generator,
    num_images_per_prompt=1,
    use_pe=True
)
    
revised_prompt = output.revised_prompts
images = output.images
image.save(f"./hf_output_0.png")
print(revised_prompt)
```