🧠 ERNIE-Image AIO for ComfyUI

All-in-One Checkpoints • BF16 & FP8 • Text Encoder + VAE Integrated • Simple ComfyUI Setup

✨ What is this?

This release provides AIO (All-in-One) checkpoint versions of ERNIE-Image and ERNIE-Image-Turbo for ComfyUI.

Instead of loading separate model parts, these AIO versions already include the image model, text encoder, and VAE inside a single checkpoint file. Just place the model into your checkpoint folder, load it with the regular Checkpoint Loader node, and build a very small workflow around it.

The original ERNIE-Image family from Baidu focuses on strong instruction following, good text rendering, structured compositions, and support for a broad range of styles, from design-oriented images to more realistic or stylized outputs.

huggingface.co Releases

• https://huggingface.co/SeeSee21/ERNIE-AIO

Workflow Releases

• https://civarchive.com/models/2553748/ernie-aio

📦 Included Versions

This release contains AIO checkpoints for:

• ERNIE-Base-AIO

• ERNIE-Turbo-AIO

Each version is provided in:

• BF16

• FP8

⚙️ AIO Format

These checkpoints are made for a simple ComfyUI workflow.

Install path

Place the .safetensors files here:

\ComfyUI\models\checkpoints

Then load them with:

• Checkpoint Loader

• CLIP Text Encode (Prompt)

• Empty Latent Image

• KSampler

• VAE Decode

• Save Image

That is all you theoretically need to run the model.

🧪 Quantization Notes

The FP8 versions in this release were quantized by me.

For these FP8 AIO builds:

• Image model: FP8

• Text encoder: FP8

• VAE: FP32

This setup was chosen to reduce size and improve practicality while keeping decoding stable and reliable.

🖼️ Recommended Usage

ERNIE-Base-AIO

• Best for: general-purpose generation, stronger prompt fidelity, more deliberate results

• Recommended CFG: 4.0

• Recommended steps: 50

ERNIE-Turbo-AIO

• Best for: fast generation, quick iteration, efficient workflows

• Recommended CFG: 1.0

• Recommended steps: 8

The official ERNIE-Image documentation recommends the same core resolutions for both releases, including:

• 1024×1024

• 848×1264

• 1264×848

• 768×1376

• 896×1200

• 1376×768

• 1200×896

🧾 Prompting

ERNIE models work best with natural language prompts.

Describe:

• subject

• scene

• composition

• lighting

• mood

• camera framing

• details you actually want in the image

These models are especially interesting for:

• posters

• text-heavy designs

• comics

• structured layouts

• prompt-sensitive image generation

🔧 ComfyUI Notes

These AIO checkpoints are designed for a checkpoint-style workflow, so you do not need to manually load separate text encoder or VAE files.

A workflow for this release will also be uploaded separately.

🙏 Credits

Huge thanks to the original ERNIE-Image team at Baidu for releasing these models and their research to the community. The original official releases are available here:

• ERNIE-Image on Hugging Face

• ERNIE-Image-Turbo on Hugging Face

• Official GitHub Repository

Both official model pages list the release under the Apache-2.0 license.

❤️ Final Note

I made these AIO versions to make ERNIE-Image easier to use in ComfyUI — especially for people who prefer the classic single-checkpoint workflow.

Load checkpoint, prompt, sample, decode, save — done.