A 3B Model Rewrites Your Prompt: The Complete Guide to ERNIE-Image PE Enhancer

ERNIE-Image's Prompt Enhancer is not a simple keyword completion tool — it's a complete 3-billion-parameter language model. Based on Mistral AI's Ministral-3B-Instruct, it is fine-tuned for one task: rewriting your prompt into a version the diffusion model can understand better.

This article fully dissects PE from a technical perspective: working principles, model architecture, parameter configuration, platform integration, and underlying behavior mechanisms.

1. Model Architecture: What is Ministral-3B?

Base Model

Property	Value
Model Name	Ministral-3B-Instruct-2512
Developer	Mistral AI
Parameters	~3B (3 billion)
Architecture	Transformer Decoder + MoE (Mixture of Experts)
Context Window	32K tokens
Release	December 2025
License	Apache 2.0 / Open-source friendly

Core Advantages of MoE Architecture

The Ministral 3 series uses a Mixture of Experts (MoE) architecture. Unlike traditional Dense Transformers that activate all parameters per inference, MoE structures activate only a subset of "expert" modules each time.

This means:

Nominal parameters: 3B
Active parameters: Significantly lower than 3B
Inference speed: Much faster than same-parameter Dense models
Resource consumption: Lower VRAM usage

For a "fast-response" preprocessing module like PE, MoE architecture is a natural advantage.

What Did Baidu Do?

Baidu performed instruction fine-tuning on top of Ministral-3B-Instruct, using high-quality paired data of "short prompt → structured long prompt".

The fine-tuned model is saved as ernie-image-prompt-enhancer.safetensors, approximately the compressed size of the base model (depending on quantization format).

2. PE's Inference Flow: Every Step from Input to Output

Step 1: Prompt Reception and Preprocessing

The user inputs an original prompt, PE receives and tokenizes it:

Raw input: "a ceramic coffee mug"
Tokenize: [a] [ceramic] [coffee] [mug]
Token count: ~4 tokens

Step 2: System Instruction Injection

PE injects a system instruction (like a Chat model's system prompt) before the user prompt, instructing the model to perform prompt expansion. Approximate system instruction:

You are a prompt enhancer for image generation. Expand the user's brief description into a detailed, structured prompt that includes subject details, environment, lighting, composition, and style. Maintain the user's original intent. Output the enhanced prompt.

Step 3: MoE Inference

Ministral-3B's MoE routing mechanism selects which expert modules to activate. For prompt expansion tasks, experts related to "text generation" and "descriptive language" are primarily activated.

Inference parameters:

max_length: 1536–2048 (output token cap)
temperature: 0.6 (balance creativity and stability)
top_p: 0.8 (nucleus sampling)
thinking mode: Disabled (skip chain of thought, output directly)

Step 4: Output Truncation and Post-processing

PE generates the expanded prompt, truncates to the specified length, then passes it to the Text Encoder (Ministral-3B Text Encoder — note this is a separate model instance, another Ministral-3B used for text encoding).

3. ComfyUI Integration Details

File Structure

ComfyUI/models/text_encoders/
├── ministral-3-3b.safetensors          ← Text Encoder (text encoding)
└── ernie-image-prompt-enhancer.safetensors  ← PE module (prompt expansion)

Note: These two models are both based on Ministral-3B but serve different purposes:

Text Encoder: Encodes prompts into text feature vectors for the diffusion model
PE module: Rewrites prompt text, output is still text

Node Configuration

PE is loaded via CLIPLoader node in ComfyUI, even though it's an LLM, not a CLIP model. This is ComfyUI's interface reuse design.

Key PE settings (via subgraph):

{
  "model": "ernie-image-prompt-enhancer.safetensors",
  "max_length": 2048,
  "temperature": 0.6,
  "top_p": 0.8,
  "thinking_enabled": false,
  "enabled": true
}

Position in Workflow

In ComfyUI's standard ERNIE-Image workflow, the PE node sits before the Text Encoder:

[CustomText (user input)] → [PE Node] → [Text Encoder (Ministral-3B)] → [DiT Diffusion Model]

If PE is disabled (enabled: false), the workflow becomes:

[CustomText (user input)] → [Text Encoder (Ministral-3B)] → [DiT Diffusion Model]

4. Diffusers API Integration

Python Code Example

from diffusers import ErnieImagePipeline
import torch
pipe = ErnieImagePipeline.from_pretrained(

"Baidu/ERNIE-Image-Turbo",

torch_dtype=torch.bfloat16

).to("cuda")
PE on (default)
image_on = pipe(

prompt="a ceramic coffee mug",

use_pe=True,          # PE enabled

height=1264,

width=848,

num_inference_steps=8,

guidance_scale=1.0

).images[0]
PE off
image_off = pipe(

prompt="a ceramic coffee mug",

use_pe=False,         # PE disabled

height=1264,

width=848,

num_inference_steps=8,

guidance_scale=1.0

).images[0]

PE's Inference Overhead

Enabling PE adds one LLM inference overhead:

Configuration	Extra Time	Extra VRAM
PE on + 8B DiT	~2-5 sec (PE inference)	~6-8 GB (PE model)
PE off + 8B DiT	0	0

In VRAM-constrained environments (12-16GB GPUs), turning off PE frees ~6-8GB for the diffusion model.

5. PE's Language Behavior Analysis

Roots of Chinese Translation Tendency

PE's tendency to translate prompts to Chinese comes from training data distribution:

Baidu's training data: Chinese prompts dominate Baidu's prompt expansion training data
Ministral base: Ministral-3B-Instruct's multilingual ability is balanced, but fine-tuning shifts toward the dominant training language
ERNIE-Image positioning: ERNIE-Image's primary target market is China, so Chinese-first is a reasonable training strategy

Impact on Multilingual Users

User Language	PE Behavior	Effect
English input	May translate to Chinese or mixed	English text rendering may become Chinese
Chinese input	Stays Chinese or mixed	Minimal impact
Other languages	May translate to Chinese	May alter original semantics

Solutions

Precise text rendering: Turn off PE
English users: Turn off PE or use external LLM
Chinese users: PE-friendly, but still turn off for text rendering

6. GGUF Quantization: PE's Blind Spot

Unsloth provides GGUF quantized versions of ERNIE-Image DiT and Ministral-3B Text Encoder, but no GGUF version exists for PE.

Likely reasons:

PE is already a 3B model — quantization benefit is limited
PE loads via CLIPLoader interface — GGUF requires a specialized Loader
Lower priority — core diffusion model and Text Encoder quantization save more VRAM

Result: GGUF workflows cannot use PE. Low-VRAM users must write detailed prompts themselves.

7. PE Parameter Tuning Guide

temperature Effects

temperature	Effect	Use Case
0.3	Conservative, stable but templated	Consistency needed in batch generation
0.6	Balanced (default)	General purpose
0.8+	Creative but may drift from original intent	Exploratory creation

max_length Effects

max_length	Effect	Use Case
512	Brief expansion	Quick preview
1536	Standard expansion (recommended)	General purpose
2048	Detailed expansion	Complex scenarios
4096+	Too long, wasteful	Not recommended

top_p Effects

top_p	Effect
0.5	Very conservative, only most likely words
0.8	Balanced (recommended)
0.95	More open, allows long-tail vocabulary

8. External LLM Alternative Implementation

Approach 1: Manual Replacement

1. Expand prompt with GPT-4/Claude
2. Turn off PE in ERNIE-Image
3. Paste expanded prompt

Approach 2: Automated Pipeline

# Pseudocode
def generate_with_custom_pe(user_prompt, llm_client):
    enhanced = llm_client.chat(
        system="Expand this image generation prompt into detailed description...",
        user=user_prompt
    )
    image = ernie_image_pipe(
        prompt=enhanced,
        use_pe=False
    )
    return image

Approach 3: Custom PE Fine-tuning

If you have extensive domain-specific prompt data, you can fine-tune PE based on Ministral-3B:

1. Download ministral/Ministral-3b-instruct
2. Prepare [short_prompt, long_prompt] paired data
3. Fine-tune with LoRA/QLoRA
4. Export as safetensors
5. Replace PE module in ComfyUI

9. Summary

PE is an instruction-fine-tuned language model based on Ministral-3B MoE architecture, whose core task is expanding short prompts into structured long prompts.

Architecture: MoE → fast inference, low VRAM
Behavior: Rewrites rather than enhances, leans toward Chinese translation
Parameters: temperature 0.6, max_length 2048, top_p 0.8 (recommended)
Limitations: Not available in GGUF workflows
Alternatives: External LLM or custom fine-tuning

Understanding PE's technical underpinnings helps you make more accurate toggle decisions and parameter optimizations.

A 3B Model Rewrites Your Prompt: The Complete Guide to ERNIE-Image PE Enhancer

Table of Contents

A 3B Model Rewrites Your Prompt: The Complete Guide to ERNIE-Image PE Enhancer

1. Model Architecture: What is Ministral-3B?

Base Model

Core Advantages of MoE Architecture

What Did Baidu Do?

2. PE's Inference Flow: Every Step from Input to Output

Step 1: Prompt Reception and Preprocessing

Step 2: System Instruction Injection

Step 3: MoE Inference

Step 4: Output Truncation and Post-processing

3. ComfyUI Integration Details

File Structure

Node Configuration

Position in Workflow

4. Diffusers API Integration

Python Code Example

PE on (default)

PE off

PE's Inference Overhead

5. PE's Language Behavior Analysis

Roots of Chinese Translation Tendency

Impact on Multilingual Users

Solutions

6. GGUF Quantization: PE's Blind Spot

7. PE Parameter Tuning Guide

temperature Effects

max_length Effects

top_p Effects

8. External LLM Alternative Implementation

Approach 1: Manual Replacement

Approach 2: Automated Pipeline

Approach 3: Custom PE Fine-tuning

9. Summary