CGFlowStudioGUIDES
Back to All Tutorials
Intermediate22 min read

Complete ERNIE-Image ComfyUI Workflow Guide: Standard vs. Turbo 8B DiT Setup & 3B Prompt Enhancer

Step-by-step tutorial on running Baidu's ERNIE-Image natively in ComfyUI — covering 8B DiT architecture, Standard (50-step) vs. Turbo (8-step), 3B Prompt Enhancer, and multilingual text rendering.

By CGFlowStudio
•
2026-08-20

Complete ERNIE-Image ComfyUI Workflow Guide: Standard vs. Turbo 8B DiT Setup & 3B Prompt Enhancer

ERNIE-Image from Baidu (released under Apache-2.0) is an open-weights 8-billion parameter Diffusion Transformer (DiT) model engineered for high-precision text rendering, complex instruction following, and structured visual layout generation.

Whether you are designing multi-panel storyboards, educational infographics, commercial poster typography in English or Chinese, or cinematic photography, ERNIE-Image offers exceptional spatial compliance and an integrated 3B Prompt Enhancer.

Available in two primary model editions — ERNIE-Image Standard (SFT) for maximum rendering quality and ERNIE-Image-Turbo for rapid 8-step generation — ERNIE-Image is now natively supported in ComfyUI. In this tutorial guide, we walk you step-by-step through installing weights, configuring the dual text encoders, utilizing the Prompt Enhancer, and mastering both Standard and Turbo workflows.

1. Introduction & Core Capabilities

Before loading your workflow, let’s explore what makes ERNIE-Image stand out in the open-source image generation landscape:

💡
Multilingual Text & Structured Layout Leadership: Unlike standard diffusion models that struggle with precise word placement, ERNIE-Image excels at rendering dense text strings, multi-panel comic layouts, and knowledge-intensive infographics in both English and Chinese.

Standard vs. Turbo Model Edition Comparison

Standard SFT ~50 Steps

ERNIE-Image (Standard)

Full 50-step Supervised Fine-Tuned (SFT) model. Delivers highest image fidelity, fine micro-textures, and complex text alignment for final production output.

Turbo Distilled 8 Steps

ERNIE-Image-Turbo

Distilled using Distribution Matching Distillation (DMD) + Reinforcement Learning (RL). Completes generation in just 8 steps for fast iteration.

2. Getting Started – Loading Official Templates

ComfyUI includes built-in native workflow templates for ERNIE-Image.

1 Update ComfyUI Core

Update ComfyUI to the latest version to register core ERNIE-Image loader nodes and prompt enhancer subgraphs.

2 Open Template Library

Click Templates in the ComfyUI top menu and search for ERNIE-Image (or Ernie Image).

3 Select Desired Workflow

Choose between ERNIE-Image (Standard) Text to Image or ERNIE-Image-Turbo Text to Image.

3. Model Installation & Directory Structure

Download and place the required 8B DiT checkpoints, Ministral 3B text encoder, 3B Prompt Enhancer, and VAE weights into your local ComfyUI subdirectories:

📂 ERNIE-Image Model Component Directory Map ComfyUI/models/
📂 ComfyUI/models/
📁 diffusion_models/
📄 ernie-image.safetensors (Standard SFT 50-Step DiT)
📄 ernie-image-turbo.safetensors (Turbo 8-Step DMD Model)
📁 text_encoders/
📄 ministral-3-3b.safetensors (Primary Ministral 3B Language Encoder)
📄 ernie-image-prompt-enhancer.safetensors (Dedicated 3B Prompt Enhancer)
📁 vae/
📄 flux2-vae.safetensors (High-Fidelity Spatial VAE)

(Official weights are available from Comfy-Org/ERNIE-Image on Hugging Face).

4. Step-by-Step Workflow Walkthroughs

Workflow 1: ERNIE-Image (Standard SFT) Mode

The Standard workflow renders images in ~50 steps for uncompromised quality and fine lighting detail.

Standard SFT Pipeline Flow
Ministral 3B Encoder → 3B Prompt Enhancer (Optional) → ERNIE-Image DiT (50 Steps) → SaveImage
  1. Load Standard Template: Open Templates -> Image -> ERNIE-Image (Standard).
  2. Verify Checkpoints: Ensure diffusion_models/ernie-image.safetensors and text_encoders/ministral-3-3b.safetensors are selected in the loader node.
  3. Enter Text Prompt: Type your target prompt into the input text node.
  4. Sampler Settings: Set steps to 50 and CFG scale to 4.0–6.0.
  5. Run Queue: Click Queue Prompt. Final rendered images save to ComfyUI/output/.

Workflow 2: ERNIE-Image-Turbo Mode

The Turbo workflow uses DMD + RL step distillation to produce final outputs in just 8 sampling steps.

Turbo Pipeline Flow
Ministral 3B Encoder → ERNIE-Image-Turbo (8 Steps) → SaveImage
  1. Load Turbo Template: Open Templates -> Image -> ERNIE-Image-Turbo.
  2. Verify Checkpoint: Ensure diffusion_models/ernie-image-turbo.safetensors is selected.
  3. Sampler Settings: Set steps to 8 and CFG scale to 1.0–2.0.
  4. Run Queue: Rapid generation completes in under 3 seconds per image.

5. Using the Built-in 3B Prompt Enhancer

ERNIE-Image includes a specialized 3-billion parameter LLM prompt enhancer (ernie-image-prompt-enhancer.safetensors) integrated directly into the conditioning stream.

Prompt Enhancer Behavior & Usage
When Enabled (`enable_enhancer = true`)

Takes short user prompts (e.g. "a cyberpunk cat") and expands them with rich lighting, background details, camera angles, and texture specifications automatically.

When Disabled (`enable_enhancer = false`)

Passes your raw text string directly to the model. Recommended when you require **strict literal prompt adherence** or exact quote typography.

6. Prompting Best Practices & Typography Syntax

ERNIE-Image excels at rendering typography, multi-subject compositions, posters, and manga panels.

Good vs. Weak Prompt Comparison

❌ Weak Prompt

"A poster with text."

Fails to specify exact text content, quote marks, font style, or spatial layout.

✅ Strong Multilingual Typography Prompt

"A vibrant retro movie poster featuring an astronaut overlooking a red planet. At the top, bold glowing neon text reads 'DESTINATION MARS'. At the bottom, golden text reads '火星探索'. Cinematic lighting, 8k resolution."

Encloses English and Chinese text inside single quotes, specifies font styling, and defines layout positions.

7. Practical Tips & Troubleshooting

💾 VRAM Considerations (~24GB Ideal)

The 8B DiT model combined with Ministral 3B and the 3B Prompt Enhancer runs natively on 24GB GPUs. For 16GB GPUs, enable CPU model offloading inside ComfyUI's model loader.

🔤 Exact Quote Syntax for Text Rendering

Always enclose exact words to be rendered inside quotes (e.g. text reading 'COFFEE TIME' or written '早安') so the text encoder handles strings as literal typography.

📂 Output File Location

All generated images are saved directly to ComfyUI/output/ with full workflow metadata embedded into the image header.

8. Conclusion & Next Steps

ERNIE-Image provides local creators with an exceptionally capable 8B DiT foundation model for multilingual text rendering and complex layout generation inside ComfyUI. By using ERNIE-Image-Turbo for fast 8-step previewing and switching to ERNIE-Image Standard for final renders, you can achieve professional graphic design outputs on your local workstation.

Next Steps to Explore

  • Test generating multi-panel comic storyboards and infographics with bilingual text overlay.
  • Compare generations with Ideogram 4.0 and Krea-2 for specialized typography projects.
  • Browse more image tutorials on our ComfyHub Tutorials Directory.