First test with flux concept.
Trained on 35mm photographs from multiple photographers mostly on landscapes, items and buildings. It took on the lighting and esthetics of the dataset pretty well but needs to be pushed into doing grainy pictures if you want those.
Use 35mm Photo, film grain in prompt for best results although it should work without.
Z-Turbo:
First real lora training for this model. This was trained using the biggest bucket available in AI-toolkit (from 768 to 1532px).
Used the same dataset as I did with Flux but added some more pictures and described the ones I had a lot better to train this one.
Whilst the model is already very capable to do photography, this has to ability to have a bit more contrast and darkness in images.
I keep the value at around 0.8
35mm is trigger word. I use high contrast or film grain / grainy shot depending on what i want to achieve as well.
Description
More steps (1536)
FAQ
Comments (8)
Thank you for putting "Z-Turbo" in the file name. I swear, most people are too lazy to name their LoRAs properly.
No problem, I too am a stickler for good naming conventions.
First, this looks great, thanks for this amazing lora! Do you have a preferred workflow for Comfy perhaps? I saw 15 steps and Euler A, but nothing more
There's nothing much more to it for the preview images, it's the most simple workflow with the model/lora's + clip going into a ksampler into an image. The only thing that sometimes differs is when I use cfg at 2 if i want more vibrancy and with Beta as scheduler . I think I the WF was downloaded from CivitAI but any basic one would do.
The results are very good and satisfying. It really improves the direct output quality and clarity of Z-Image Turbo. I can say that the training with 1536 pixels yielded very good results. Would you consider training for Wan and Qwen as well?
I'd be interested in WAN since I use that one frequently as well. But since training on 1536px takes ages, I'm not sure it'll be anytime soon. The time needed really increases from 1024 or 1280px.
When i combine this lora with my character lora the image ends ruined :(
What weights are you using for the Lora's? I've noticed Z-Turbo doesnt mix too great with different lora's and I need to keep other lora's to a max of 0.2-0.4 if I use a character to keep the likeness
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.