No trigger word needed for this LoRA. It should improve the quality of images generated for anime and digital illustration overall.
You should pair it with another LoRA to steer the model towards the style you want your image to have, this one should only aims to increase quality.
The images in the preview are NOT quality maxxing (you can do better with precise style LoRA & up-scaling), it just showcases what the LoRA outputs.
Z-Image (Base):
It happears that LoRA trained on base seem to work on Turbo if you crank the strength very high (between 2 and 4)
Honesty, Z-Image is such a good base you might even need a LoRA to steer it towards high quality illustration. It is the default. The LoRA has only a very light effect, especially if you already schizo-prompt for a specific style.
This model can pretty much do anything with the right prompting, though it is a bit slow (Chroma like speed).
Really hyped to see the fine-tune that will come out of this because the base model is already able to do some nsfw (Tongyi-MAI must have trained quite a bit on NSFW judging by how it does genitals / nsfw scenes, then abliterated before release)
Klein:
Klein (both 4b and 9b) are very schizo model, they are extremely good a certain things (landscape, geometry, high-res) and worse than flux 1 at other (style, text, spacial relations between objects).
They will 80% of the time ignore half of your prompt while generating the worse body horror since SD1.4. Then generate a nearly perfect image the next seed.
That being said, I think klein 4b will be the next SDXL. It is definitely a step up from Illustrious in some way for anime (and other styles). While still being faster than ZImage and it has edit abilities.
For the time being both klein model are worse than Chroma if you just want to normal photo / anime images. I'll use Klein 4b and Z-Image as an upscale model for Chroma until good fine-tune come out. I invite you to give Chroma a shot, but keep in mind that the better instruction following and style of chroma comes with the cost of it being very slow. Read this article before starting to use it.
Description
Trained on 3500 hand picked anime / digital illustration from civitai.
Onetrainer was used with Adamw at batch size 4 for 8600 steps at 512, 1024px
FAQ
Comments (5)
Is this the first lora I see there for Z-Image Base or it is still Z-Image Turbo. I know we got no catgory for base right now, but this is in category Z-Image-Turbo. Please quick clarification. Thank you. :)
This is trained on the base model in OneTrainer for ~2.5k steps bs8 at 512px with the Muon optimizer. (It might be under-trained since this is my first Z-Image lora)
I've seen a few other lora on discord and on civit. You can train them in both OneTrainer and Ai-Toolkit
Z-Image base seem to be compatible with ZIT if you put them at really high weight (This generates digital illustration if you set the strength to 3 on ZIT)
@agile363274 So far this is the only lora ZiB on CivitAI and it is in category ZiT, can you post any other lora for ZiB, I was looking everywhere on CivitAi. What filters do you have? (can confirm no other ZiB Loars at this time for me)
Also thanks for tip, I only heard that they are not compatible, but can try later with much higher strength.
@HugMeIntoFace Z-Image-Aesthetic was posted with a base version before.
From my few test with this LoRA, they seem to be at least mostly compatible when it comes to style while cranking the strength up. Though time will tell on what model you need to train to get the best result for Turbo.
I know I will only use base since the style of Turbo annoys me quite a bit, but I hope people can use my LoRAs with it too.
@agile363274 Thank you, i had it in wrong folder. Still first for style. Thank you again, will look forward for more. ;)
Yeah turbo is more for realism. The first look on the base it was looking that it was more distinguishes between styles. I presume that now begins the era of Z-Image. Looking forward to it.
GL with more stuff!!!!