[Feature] ERNIE-Image

[Green-Sky]

Feature Summary

Looks like a new good model just dropped.

Detailed Description

https://huggingface.co/baidu/ERNIE-Image
https://huggingface.co/baidu/ERNIE-Image-Turbo
https://huggingface.co/Comfy-Org/ERNIE-Image

ERNIE-Image is an open text-to-image generation model developed by the ERNIE-Image team at Baidu. It is built on a single-stream Diffusion Transformer (DiT) and paired with a lightweight Prompt Enhancer that expands brief user inputs into richer structured descriptions. With only 8B DiT parameters, it reaches state-of-the-art performance among open-weight text-to-image models. The model is designed not only for strong visual quality, but also for controllability in practical generation scenarios where accurate content realization matters as much as aesthetics. In particular, ERNIE-Image performs strongly on complex instruction following, text rendering, and structured image generation, making it well suited for commercial posters, comics, multi-panel layouts, and other content creation tasks that require both visual quality and precise control. It also supports a broad range of visual styles, including realistic photography, design-oriented imagery, and more stylized aesthetic outputs.

Seems to be similar to z-image, but with flux2-vae.

[JohnLoveJoy]

It's impressive. Better than both Klein 9B and Qwen Image.

[Green-Sky]

There seems to be a prompt enhancer "pe" too.
https://huggingface.co/baidu/ERNIE-Image-Turbo/tree/main/pe

[JohnLoveJoy]

There seems to be a prompt enhancer "pe" too. https://huggingface.co/baidu/ERNIE-Image-Turbo/tree/main/pe

It's a bit more complex than that.

[stduhpf]

VAE from Flux2
Text encoder seems to be Ministral 4B (Mistral3 architecture with ~4B parameters), like the "pe" model

DiT seems to be a fairly simple single stream architecture

[leejet]

#1427