ERNIE-Image

ERNIE-Image ComfyUI Official Workflow Deep Dive: From Zero Configuration to Production Pipeline

Jun 13, 2026

ERNIE-Image ComfyUI Official Workflow Deep Dive: From Zero Configuration to Production Pipeline

Abstract: ComfyUI now provides complete official workflow templates for ERNIE-Image, supporting both Base (50 steps for quality) and Turbo (8 steps for speed) pipelines. This article starts from scratch, covering the official workflow's node chain, model file configuration, Base/Turbo differences, multi-model switching strategies, and advanced techniques from single-image generation to batch production. Combined with practical insights from the 15K+ view YouTube tutorial series, this guide takes you from beginner to production-level ComfyUI ERNIE-Image user.

The Milestone of Official ComfyUI Support

In early 2026, ComfyUI officially announced full support for ERNIE-Image. This is not community adaptation — it's official template system level integration:

  1. Native Integration: ComfyUI v0.3.0+ ships with ERNIE-Image workflow templates built-in
  2. One-click Loading: Templates → ERNIE-Image directly imports complete workflows
  3. Auto Model Download: Required model files fetched automatically via HuggingFace
  4. Dual Model Support: ERNIE-Image Base (50 steps, high quality) and ERNIE-Image-Turbo (8 steps, fast iteration)

This marks ERNIE-Image's transition from "community manual adaptation" to "official ecosystem support."

Core Architecture: ERNIE-Image Node Chain in ComfyUI

The ERNIE-Image ComfyUI workflow consists of these core nodes:

[CLIP Text Encode] ──→ [ERNIE-Image Diffusion Model] ──→ [VAE Decode] ──→ [Save Image]
        ↑                        ↑                                ↑
   Prompt Text            ernie-image.safetensors          flux2-vae.safetensors
   (or PE output)            (or turbo.safetensors)

Component Breakdown:

Component File Role Source
Diffusion Model (Base) ernie-image.safetensors Main generation model (8B DiT) Comfy-Org/ERNIE-Image
Diffusion Model (Turbo) ernie-image-turbo.safetensors Distilled fast version (8 steps) Comfy-Org/ERNIE-Image
Text Encoder ministral-3-3b.safetensors Ministral 3B text encoder baidu/ERNIE-Image
Prompt Enhancer ernie-image-prompt-enhancer.safetensors 3B prompt enhancer baidu/ERNIE-Image
VAE flux2-vae.safetensors FLUX.2 Variational Autoencoder FLUX.2

⚠️ Important: ERNIE-Image uses FLUX.2's VAE (flux2-vae.safetensors), NOT traditional SD VAE. Using the wrong VAE is the #1 cause of blurry or color-distorted images in ComfyUI ERNIE-Image workflows.

Quick Start: From Installation to First Image

Step 1: Update ComfyUI

cd ComfyUI
git pull origin master
# Ensure version ≥ v0.3.0

Step 2: Load ERNIE-Image Workflow

  1. Open ComfyUI web interface (default http://127.0.0.1:8188)
  2. Click Templates tab
  3. Search for ERNIE-Image
  4. Select ERNIE-Image (Base) or ERNIE-Image-Turbo workflow
  5. Click Load

Step 3: Download Model Files

ComfyUI auto-detects missing models and prompts download. Place files in:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── ernie-image.safetensors          # Base (~16 GB)
    │   └── ernie-image-turbo.safetensors    # Turbo (~16 GB)
    ├── text_encoders/
    │   ├── ministral-3-3b.safetensors       # Text encoder (~6 GB)
    │   └── ernie-image-prompt-enhancer.safetensors  # PE
    └── vae/
        └── flux2-vae.safetensors            # FLUX.2 VAE (~2 GB)

Disk space needed: ~30-35 GB for full installation.

Step 4: Generate Your First Image

  1. Enter a prompt in CLIP Text Encode node
  2. Set sampler (Turbo: Euler, 8 steps; Base: DPM++ 2M Karras, 20-50 steps)
  3. Set CFG Scale (Turbo: 1.0, Base: 5.0-7.0)
  4. Click Queue Prompt to start

Base vs Turbo: Workflow Comparison

Feature Base Turbo
Parameters 8B DiT 8B DiT (DMD+RL distilled)
Steps 20-50 8 (fixed)
CFG Scale 5.0-7.0 1.0 (fixed)
Sampler DPM++ 2M Karras Euler
Quality Best Near-Base
Speed Slower 3-6× faster
Best For Final output, print Quick iteration, preview

Key difference: Turbo uses the same text encoder and VAE — just swap the diffusion model file. In ComfyUI, switching models means changing the Load Diffusion Model node input.

💡 Practical tip: Use Turbo during development to quickly validate prompts and parameters, then switch to Base for final output. This "dual pipeline" strategy is recommended by Pixaroma in the YouTube Ep14 tutorial (15,898 views).

Practical Test Scenarios: Community Benchmarks

Based on YouTube Ep14 tutorial (15,898 views) and community feedback:

Scenario ERNIE-Image Base ERNIE-Image Turbo
Food Photography ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐
Text Rendering ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐
Anime Style ⭐⭐⭐⭐ ⭐⭐⭐⭐
Realistic Portrait ⭐⭐⭐⭐ ⭐⭐⭐⭐
UI/Interface Design ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐
Painting Style ⭐⭐⭐⭐ ⭐⭐⭐⭐

Vs Competitors (same 8-step Turbo scenario):

  • vs Z-Image Turbo: ERNIE-Image leads in text rendering
  • vs Nano Banana 2: ERNIE-Image more stable in instruction following
  • vs ChatGPT Images 2: ERNIE-Image supports local deployment and LoRA fine-tuning

Advanced Techniques: From Single Image to Production Pipeline

Technique 1: Multi-Model Switching

Support both Base and Turbo in the same workflow:

[Switch Model] → [Load Diffusion Model (Base or Turbo)]
       ↑
  [Boolean Toggle Node: Switch Base/Turbo]

Use ComfyUI's Primitive Node or conditional routing for one-click model switching.

Technique 2: Batch Prompt Input

For e-commerce product photos, social media batch generation:

[Load Text (prompts.txt)] → [Batch Prompt Split] → [CLIP Text Encode] → [KSampler] → [Save Image (sequential)]

Each line in prompts.txt is one prompt. ComfyUI auto-saves outputs with sequential numbering.

Technique 3: VRAM Optimization

For hardware under 24 GB VRAM:

  1. FP8 Quantization: Use FP8 diffusion model, ~50% less VRAM
  2. --lowvram flag: Start ComfyUI with --lowvram
  3. Staged generation: Generate at 512×512, then upscale to 1024×1024
  4. Turbo-first: Always use Turbo during iteration

Technique 4: Custom Node Extensions

Node Function Source
Pixaroma Note Rich text editor with color coding, code blocks, lists Pixaroma
Resolution Node Smart resolution calculation, preset sizes Same
Wildcard Randomized prompt variables for diversity ComfyUI-Wildcards
Aspect Ratio Preset aspect ratio selection ComfyUI-Aspect-Ratio

Install custom nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/pixaroma/ComfyUI-Pixaroma
pip install -r requirements.txt

FAQ

Q: Nodes report errors after loading the template?

Update ComfyUI to the latest version and check ComfyUI/custom_nodes for missing dependencies. Run pip install -r requirements.txt.

Q: Not enough VRAM?

  1. Use the Turbo model
  2. Enable --lowvram startup parameter
  3. Use FP8 quantized model
  4. Lower output resolution then upscale

Q: Generated text is unclear?

  1. Enable Prompt Enhancer (significantly improves text rendering)
  2. Increase sampling steps (30-50 for Base)
  3. Increase CFG Scale (7.0-8.0)

Q: Are Base and Turbo LoRA weights compatible?

Yes — Base and Turbo share the same architectural foundation, so LoRA weights are generally interchangeable. Turbo's distillation may cause slight differences in some LoRA effects.

Summary

ComfyUI's official support for ERNIE-Image marks the maturation of the open-source text-to-image workflow ecosystem. From basic templates to custom node extensions, from single-image generation to batch production, ERNIE-Image's ComfyUI integration is now production-ready.

Recommended workflow strategy:

  1. Beginners: Use official templates, start with Turbo
  2. Advanced: Combine Pixaroma Note + PE enhancement
  3. Production: Batch workflows + multi-model switching
  4. Optimization: Turbo iteration + Base output dual-pipeline

Keywords: ernie-image comfyui official workflow ernie-image comfyui template ernie-image comfyui base turbo ernie-image production pipeline ernie-image comfyui tutorial comfyui ernie-image model setup

ERNIE-Image Team