ERNIE-Image Prompt Weighting and Negative Prompt Advanced Guide: Complete CFG Scale / Negative Prompt Analysis

Published: 2026-06-05
Author: ERNIE-Image Technical Team
Keywords: ernie-image cfg scale ernie-image negative prompt ernie-image guidance scale ernie-image prompt weighting

---

Introduction

In AI image generation, the prompt is the core mechanism for controlling output quality. But for ERNIE-Image—a Diffusion Transformer model—having a good prompt text alone isn't enough. You need to master three key technical parameters: CFG Scale (Guidance Scale), Negative Prompt, and Prompt Weighting to achieve precise control over generation results.

Many new users encounter these frustrations with ERNIE-Image:

• "Why does my negative prompt have no effect at all?"
• "Why does raising CFG to 7 make the quality worse?"
• "How do I make the model focus more on a specific element in my prompt?"

This article answers each question and provides verified best practices.

---

1. CFG Scale (Guidance Scale): Controlling Prompt Adherence

1.1 What is CFG Scale?

CFG (Classifier-Free Guidance) Scale controls how closely the model follows the text prompt during generation. Simply put:

• CFG = 1: Model almost ignores the prompt, free-form generation
• CFG = 4.0 (default): Balanced adherence and creative freedom
• CFG ≥ 7: Strictly follows the prompt, but may lose color naturalness and diversity

1.2 ERNIE-Image Recommended CFG Values

Mode	Default	Recommended Range	Notes
ERNIE-Image (Base)	4.0	3.0 - 5.0	Standard generation
ERNIE-Image-Turbo	1.0	1.0 - 2.0	DMD+RL optimized, low CFG works
Upscaling Phase	3.0	2.5 - 3.5	Lower CFG preserves original structure
Text Rendering	4.5	4.0 - 5.5	Higher CFG enhances text accuracy

Key difference: ERNIE-Image-Turbo is optimized with DMD and RL, and achieves high quality at the default CFG of 1.0. Setting Turbo's CFG above 4.0 can actually produce oversaturation and artifacts.

1.3 CFG Scale Impact on Output Quality

CFG 1.0  →  Soft colors, free composition, may drift from prompt
CFG 2.5  →  Balanced state, the sweet spot for most scenarios
CFG 4.0  →  Strict prompt adherence, standard recommended value
CFG 5.0  →  Highly adherent, precise details but may be oversaturated
CFG 7.0+ →  Color distortion, possible over-sharpening or artifacts

1.4 CFG in Diffusers

from diffusers import ERNIEImagePipeline
pipe = ERNIEImagePipeline.from_pretrained("baidu/ERNIE-Image")
Standard generation
image = pipe(

prompt="a cat sitting on a windowsill",

guidance_scale=4.0,  # Default recommended value

num_inference_steps=50,

).images[0]
Turbo mode — low CFG
image_turbo = pipe(

prompt="a cat sitting on a windowsill",

guidance_scale=1.0,  # Turbo default

num_inference_steps=8,

).images[0]

---

2. Negative Prompt: Excluding Undesired Content

2.1 Does ERNIE-Image Support Negative Prompts?

Yes, but with conditions. ERNIE-Image is based on the DiT architecture, and its Diffusers pipeline accepts the negative_prompt parameter. However, DiT models respond differently to negative prompts compared to traditional SD 1.x/2.x:

• Requires higher CFG Scale: Negative prompt effects are noticeable at CFG ≥ 3.0
• Limited effect in Turbo mode: Turbo's default CFG of 1.0 makes negative prompts nearly ineffective
• Separate Negative Guidance Scale in ComfyUI: Can be adjusted independently

2.2 Negative Prompt Best Practices

from diffusers import ERNIEImagePipeline
pipe = ERNIEImagePipeline.from_pretrained("baidu/ERNIE-Image")
Effective use of negative prompt
image = pipe(

prompt="a professional headshot portrait of a woman, studio lighting",

negative_prompt="cartoon, anime, drawing, painting, blurry, low quality, deformed, ugly, bad anatomy, extra limbs",

guidance_scale=4.5,  # Higher CFG needed

num_inference_steps=50,

).images[0]

2.3 Negative Prompt Considerations

Scenario	Recommendation
Base model + CFG ≥ 3.0	✅ Negative prompt effective
Turbo model	⚠️ Limited effect (CFG defaults to 1.0)
In ComfyUI	✅ Adjustable via Negative Guidance Scale
Text rendering	⚠️ May interfere with text generation, use sparingly

Important: ERNIE-Image-Turbo's DMD-distilled optimization uses default CFG=1.0, making negative prompts nearly ineffective. If you need negative prompts, use the Base model or raise Turbo's CFG to 3.0+ (but this may compromise Turbo's quality advantage).

---

3. Prompt Weighting: Fine-Controlling Element Importance

3.1 Why Do We Need Prompt Weighting?

In practice, you often need to emphasize certain elements in your prompt while deemphasizing others. For example:

• "A very detailed cat on a windowsill" → Want the model to focus more on cat details
• "Cyberpunk style city with pedestrians" → Want style to be more important than content

ERNIE-Image supports multiple prompt weighting syntaxes.

3.2 Prompt Weighting Syntax

Method 1: Parenthesis Weighting (word)

Add parentheses around words/phrases you want to emphasize:

# No weighting
"a cat on a windowsill with flowers"
Weighted — cat is more important
"(cat:1.3) on a windowsill with flowers"
Multiple weights — cat is most important, flowers secondary
"(cat:1.5) on a windowsill with (flowers:1.2)"

Method 2: Repetition Emphasis

Repeat keywords to increase their importance:

# Moderate emphasis
"cat, cat, on a windowsill"
Strong emphasis
"cat, cat, cat, on a windowsill"

Method 3: Position Priority

Diffusion models pay more attention to words at the beginning of the prompt. Place the most important descriptions first:

# Cat is the focus
"close-up portrait of a cat, detailed fur, sitting on windowsill, flowers in background"
Flowers as focus (not recommended, poor effect)
"sitting on windowsill, flowers, close-up portrait of a cat, detailed fur"

3.3 Weight Coefficient Reference

Weight	Effect	Use Case
1.0	Standard weight	Default, no emphasis needed
1.2	Slight emphasis	Secondary elements need a bit more attention
1.3-1.5	Moderate emphasis	Core elements need more attention
1.5+	Strong emphasis	Key elements that must stand out
0.5-0.8	Reduced weight	Elements to de-emphasize

Note: Excessively high weights (>2.0) can cause attention imbalance, producing anomalous outputs. Recommended weight range: 0.5-1.5.

---

4. Advanced Control in ComfyUI

4.1 CFG and Negative Guidance Scale in ComfyUI

ERNIE-Image workflows in ComfyUI provide finer-grained control:

KSampler Node:
├── positive — positive prompt
├── negative — negative prompt
├── cfg — CFG Scale (global)
└── denoise — denoising strength (img2img mode)

For Flux/DiT-class models, ComfyUI supports Negative Guidance Scale, which can be set independently in the CLIPTextEncodeFlux node:

CLIPTextEncodeFlux Node:
├── text — prompt text
├── guidance — this prompt's weight (default 3.5)
└── (negative guidance scale — negative weight)

4.2 Advanced Negative Prompt Usage in ComfyUI

The ComfyUI community discovered an interesting technique: even with an empty negative prompt, raising the Negative Guidance Scale can affect output quality:

Positive prompt: "a beautiful landscape"
Negative prompt: "" (empty)
CFG Scale: 6.0
Negative Guidance Scale: 30 (high value increases contrast and sharpness)

This technique is widely verified on FLUX Dev models, and ERNIE-Image as a same-architecture model may also benefit.

---

5. Best Parameter Combinations for Different Scenarios

5.1 Quick Reference Table

Scenario	Model	CFG	Negative Prompt	Steps
Quick concept validation	Turbo	1.0	Not recommended	8
Standard photography	Base	4.0	Generic negative words	50
Text rendering	Base	4.5	Minimal/none	50
Anime style	Base	3.5	"realistic, photo"	50
Product photography	Base	4.0	"blurry, low quality"	50
Upscaling refinement	Base	3.0	Minimal	20-30
Stylized creation	Base	2.5-3.0	Style-dependent	50

5.2 Universal Negative Prompt Templates

# Universal quality negatives (for most scenarios)
negative_prompt = "blurry, low quality, worst quality, deformed, distorted, disfigured, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, mutation, watermark, text, signature, username"
Photorealistic scenarios
negative_prompt = "cartoon, anime, drawing, painting, illustration, 3d render, blurry, low quality, deformed"
Anime/illustration scenarios
negative_prompt = "photorealistic, realistic photo, blurry, low quality, deformed, watermark"
Text rendering scenarios (use carefully)
negative_prompt = "blurry text, illegible text, distorted text"

---

6. FAQ and Troubleshooting

Q1: Negative prompt has no effect at all?

Troubleshooting steps:

1. Confirm you're using the Base model (not Turbo)
2. Confirm CFG Scale ≥ 3.0
3. In ComfyUI, confirm negative prompt is correctly connected to KSampler
4. Try simplifying the negative prompt (3-5 core words is enough)

Q2: Image looks ugly/oversaturated when I raise CFG?

• Lower CFG to 3.0-4.0 range
• Reduce inference steps (from 50 to 30)
• Try adding "natural lighting, soft colors" to your prompt

Q3: How do I make a specific element stand out?

• Use parenthesis weighting: (element:1.5)
• Place that element at the beginning of the prompt
• Add competing elements to negative prompt: e.g., "a cat, not a dog"

Q4: Negative prompt ineffective in Turbo mode?

This is normal behavior. Turbo's default CFG=1.0 makes negative prompts weak. Solutions:

• Use Base model for generation, then Turbo for upscaling/refinement
• Or manually raise Turbo's CFG to 3.0 (at the cost of some speed)

---

7. Summary

Mastering CFG Scale, Negative Prompt, and Prompt Weighting is the key to unlocking ERNIE-Image's full potential:

1. CFG Scale: Base model recommended 3.0-5.0, Turbo recommended 1.0-2.0. Too high causes oversaturation; too low drifts from the prompt.
2. Negative Prompt: Effective with Base model + CFG ≥ 3.0. Limited effect in Turbo mode.
3. Prompt Weighting: Use parenthesis weighting (word:1.3) and position-priority strategies for fine-grained element control.

Using these three parameters together, you can evolve from "luck-based generation" to "precise output control."