Video - https://youtu.be/W09xam6Ly4k

Overview

Ernie Image is now supported in the very latest version of ComfyUI and it is surprisingly good!

There are both turbo and non-turbo versions available, and their results tables look pretty competitive over at https://huggingface.co/baidu/ERNIE-Image

Is it really better than Z-Image-Turbo & Qwen-Image? Yes... in most ways, at least from the tests I've done so far.

As well as the basic Text-To-Image tasks, ERNIE-Image also does rather well with Image-2-Image in In-Painting tasks, albeit with slightly higher denoise values than other models.

The range of styles is impressive, and the ability to actually follow prompt instructions is very high indeed. There does seem to be quite the Chinese bias / default when it comes to the appearance of humans (and dragons!), but you can simply prompt your way out of it by being more descriptive. It did the English village scene above very well, for example!

Workflow Information

There are 2 workflow versions attached - one for ERNIE-Image only, and one with both Z-Image & ERNIE-Image as well. As described in the note, you may wish to try one model or the other first and then upscale with something else:

If both are enabled, the default will be Z-Image (Turbo). *_E = Ernie, *_Z = Z-Image. By default it is set up to just use one model at a time, allowing for easy switching - great when seeing which model actually works best in a given situation!

As usual, model download links are available in-workflow:

So far, I've mostly been testing and comparing the two Turbo versions:

But the workflow has also be used with the non-turbo models as well - just change the CFG and Steps!

Whilst Ernie Image is limited to around 1k resolution (compared to 2k for ZIT), it does seem to have a much better grasp on styles and prompt adherence.

Upscaling does mean Ernie can also reach 2k (or more, with ultimate upscale), though the Turbo version does seem to converge much fast than other models. This means the split-sigmas only needs 2-3 steps before really locking-in the image. When using anything with a sub-1 denoise value, it tends to need a little more noise than other models (such as around 0.8 compared to 0.5, assuming the simple scheduler).

ERNIE-Image also comes with a Text Prompt Enhancer, which also works great with ZIT as well.

If you are looking to use one model and then the other, enable the second prompt...

... and even set a different negative prompt, if you like:

With both models loaded, the default is Z-Image. Thus, to start with Z-Image first and then use ERNIE for 2nd stage tasks, you can simply select the ERNIE models in Stage 2:

This will tend to give you a nice upscale.

Example Image 1 - Z-Image-Turbo:

Image 2 - ERNIE-Image Upscale:

Image compare (Ernie on the left, Z-Image on the right):

You could, of course, start with ERNIE instead and then upscale with Z-Image - whatever you prefer! Just remember to set the MODEL, Prompts & VAE values (like in the screenshot above) in whichever group you don't want to use Z-Image (Turbo). If using Non-Turbo models in Stage 2, also remember to change the CFG. No "latent interposer" nodes yet for these, so only really works with Image Upscale groups (IMG Upscale & Ultimate Upscale).

Most options are turned OFF by default:

Simply flip the relevant switch to enable whatever it is you'd like to try! Clicking the arrow will take you straight to that group. The workflows also have numbered bookmarks (1-7) for fast navigation (1 = load -> 7 = Image Compare).

The Image/Latent GET/SET nodes can also be changed to whatever values you like, such as in the compare image nodes:

Thus, as an example, if you're looking to compare your Image Upscale and Ultimate Upscale group results, that is super easy to do just by changing the GET node value.

Enjoy! :)