SeeSee21 / ERNIE-AIO

ERNIE AIO for ComfyUI

All-in-One Checkpoints • Base + Turbo • BF16 & FP8 • Integrated Text Encoder + VAE

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

A ready-to-use ComfyUI workflow is included in the Workflow/ folder, and the optional Prompt Enhancer files are included in the prompt-enhancer/ folder.

These AIO releases are designed for users who prefer a classic single-checkpoint workflow .
Each checkpoint already contains the image model , text encoder , and VAE , so no separate text encoder or VAE loading is required.

ERNIE-AIO-Base

ERNIE-AIO-Turbo

Overview

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

These AIO releases are designed for users who prefer a classic single-checkpoint workflow .
Each checkpoint already contains the image model , text encoder , and VAE , so no separate text encoder or VAE loading is required.

Available Files

Main checkpoints

• ERNIE-AIO-Base-bf16.safetensors
• ERNIE-AIO-Base-fp8.safetensors
• ERNIE-AIO-Turbo-bf16.safetensors
• ERNIE-AIO-Turbo-fp8.safetensors

Workflow

• Workflow/ERNIE-AIO-Upscale.json

Prompt Enhancer

• prompt-enhancer/ contains the optional ERNIE Prompt Enhancer files used by the workflow

What “AIO” means here

The AIO checkpoints are packaged so they can be placed directly in:

ComfyUI/models/checkpoints

and loaded with the normal Checkpoint Loader node in ComfyUI.

This means you can use a very small workflow such as:

1. Checkpoint Loader
2. CLIP Text Encode (Prompt)
3. Empty Latent Image
4. KSampler
5. VAE Decode
6. Save Image

A workflow file may also be included separately for convenience.

Model Background

These AIO releases are based on Baidu’s official ERNIE-Image family.

According to the official model pages:

• ERNIE-Image is an open text-to-image model built on a single-stream Diffusion Transformer (DiT) and paired with a lightweight Prompt Enhancer
• ERNIE-Image-Turbo is the distilled variant, designed for strong fidelity in only 8 inference steps
• The official releases emphasize instruction following , text rendering , and structured image generation , making them especially interesting for posters, comics, infographics, UI-like layouts, and other text-heavy or layout-sensitive images

AIO Variants

ERNIE-AIO-Base

The Base version is intended for more deliberate, stable, and higher-quality generation.

Recommended settings:

• Steps: 50
• CFG: 4.0

Use this version when you want stronger refinement, more careful rendering, and the full non-turbo experience.

ERNIE-AIO-Turbo

The Turbo version is intended for fast generation and quick iteration.

Recommended settings:

• Steps: 8
• CFG: 1.0

Use this version when you want faster turnaround while still keeping good prompt fidelity and strong visual output.

Precision / Quantization Notes

The FP8 versions in this repository were quantized by me.

For these FP8 AIO checkpoints:

• Image model: FP8
• Text encoder: FP8
• VAE: FP32

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

Prompt Enhancer

This repository also includes an optional Prompt Enhancer in the prompt-enhancer/ folder.

It can be used with the included workflow to automatically expand short prompts into richer and more descriptive prompts.
This is optional, but useful if you want more detailed prompt wording or a more guided workflow experience.

Prompting

These models work well with natural language prompts .

They are especially suitable for:

• posters
• title cards
• product ads
• readable text in images
• structured layouts
• comics
• infographics
• UI-like visuals
• realistic and stylized imagery

For best results, describe:

• subject
• environment
• composition
• mood
• lighting
• camera framing
• text that should appear in the image

Example Prompt Ideas

• A cinematic movie poster with a bold readable title at the top, dramatic lighting, and a futuristic skyline.
• A modern coffee advertisement with readable packaging text and a premium warm studio look.
• A clean infographic with clearly labeled sections and a structured layout.
• A comic-style scene with readable speech bubbles and strong composition.

ComfyUI Usage

Place the checkpoints in:

ComfyUI/models/checkpoints

Then load them through Checkpoint Loader like a normal checkpoint.

No separate:

• VAE loader
• text encoder loader
• extra model assembly

is required for the AIO versions.

Repository Structure

• Main AIO checkpoints for ERNIE-AIO-Base and ERNIE-AIO-Turbo
• Workflow/ contains the ready-to-use ComfyUI workflow JSON
• prompt-enhancer/ contains the optional Prompt Enhancer files used by the workflow
• images/ contains the preview / title images for the Base and Turbo releases

Civitai Releases

If you prefer the Civitai pages, you can also find the releases here:

• Turbo: https://civitai.red/models/2551548?modelVersionId=2868512
• Base: https://civitai.red/models/2551548?modelVersionId=2867462

Limitations

As with other image generation models:

• text rendering can still vary depending on prompt complexity
• highly dense typography may require retries
• prompt phrasing can strongly affect layout quality
• FP8 variants may behave slightly differently from BF16 variants depending on hardware and workflow

Intended Use

These releases are intended for:

• local image generation in ComfyUI
• poster and design-style generation
• readable title images
• structured compositions
• users who want a simple all-in-one checkpoint workflow

Credits

Huge thanks to the original ERNIE-Image team at Baidu for releasing the official models.

Official sources:

• ERNIE-Image: https://huggingface.co/baidu/ERNIE-Image
• ERNIE-Image-Turbo: https://huggingface.co/baidu/ERNIE-Image-Turbo
• GitHub: https://github.com/baidu/ERNIE-Image

License

This repository is based on the official ERNIE-Image releases, whose Hugging Face model pages list the license as Apache-2.0 .

Final Note

These AIO releases were made to simplify the ERNIE workflow for ComfyUI users.

Load checkpoint, write prompt, sample, decode, save.