Main characteristics of this LoRA are as follows:
· Reversal film / slide film style, with strong contrast and rich, intense color saturation.
· Improved lighting and shadow effects, especially extremely powerful and expressive sunlight rendering.
· Based on personal aesthetic preference, female characters tend toward large breasts, slim waist, and wide hips.
Important: This LoRA has a very strong style with distinct personal preferences. In other words, it cannot guarantee diversity in body types etc. (You can achieve more variety by using detailed and specific prompts.)
Info for ERNIE-image (and turbo) v1.1
This LoRA is compatible with ERNIE-image and its distilled model ERNIE-image-turbo.
For the ERNIE-image model, the default recommended settings are CFG 4.0 / steps 25+.
It can also be used in combination with the Turbo LoRA (GuangyuanSD/ErnieTurbo_extracted_lora · Hugging Face) at CFG 1.0 / steps 8.
For the ERNIE-image-turbo model, it can be used directly.
In addition, this LoRA can significantly reduce the yellowish cast and rough artifacts of the ERNIE-image-turbo model, producing images with better white balance and finer, more detailed quality.
Recommended strength: 1.0
Info for z-image-turbo v2.0
This is an improved version of v1.0, with the main changes as follows:
· Overall tone is brighter, with more details preserved in the shadows
· Reduced some unreasonable light and shadow effects
· Contrast is not as strong/intense as in v1.0
Rather than calling it an “upgrade,” it’s more accurate to say it’s a variant with a different color tone and style. Users can choose between the two versions based on their personal preference.
Recommended strength: 0.7–1.0
Recommended sampler & scheduler: euler_capitanZIT (https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler)
Info for z-image-turbo v1.0
Recommended strength: 0.5–1.0
Recommended sampler & scheduler: euler_capitanZIT (https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler)
Info for z-image v1.0
This is a version retrained based on the z-image model.
Recommended strength: 1.0
Recommended sampler & scheduler: res_multistep beta 25steps
Description
FAQ
Comments (10)
Thanks I'm getting very beautiful results with this.
Are you planning to make a base Z-image lora for this one? It's my favourite Z-image lora so far lol
I am indeed currently training the z-image version.
Unfortunately, LoRAs trained for z-image are not compatible with z-image-turbo, which is quite a letdown for me.
From my perspective, turbo already satisfies all my requirements with sufficient performance, so I likely won't invest much more time or effort into z-image in the future.
That said, I plan to release a z-image version of this LoRA anyway—it should take a few days.
@AIImageStudio Ah I didn't know that, yeah that's sad, and suprising, I thought that's why people were excited for base, to get beter loras for turbo.
One thing I do like about Z-image base is the variety it gives you, look at this thread to give you a better idea. Also listens to the prompt way more e.g mole location, different looking women together, Japanese woman actually looking Japanese etc. there's an image in this thread that shows that too, also negative prompts are working now. So for me that's important, although with loras might not matter as much. But yeah slower for sure.
https://www.reddit.com/r/StableDiffusion/comments/1qozyms/a_quick_test_showing_the_image_variety_of_zimage/
But yeah thanks, I look forward to your lora!
@Light6969 I'm kinda stuck here... The LoRA training for z-image isn't going anywhere no matter what I do. Even if I increase the steps a lot, there's basically no visible change. Using the exact same dataset and parameters, I still can't get anywhere close to the quality of z-image-turbo...
@AIImageStudio Damn, yeah early days, might take time for people to figure out the best settings for base. Anyway if the worst comes to the worst, I can still use your turbo lora in z-image turbo! Don't worry about me, I don't wanna give you extra pressure haha.
@Light6969
Thank you.
I might release a preview version later today. This version has some effect, but it’s not as good as z-image-turbo.
After that, I’ll keep an eye on other people’s progress and see if I need to adjust the training approach.
@AIImageStudio Ah cool! Yeah I'll keep an eye out too, maybe I'll get into lora training one day too.
@Light6969 preview version released
@AIImageStudio Nice I'll test it out later tonight. :D