AfterDark for Z-Image Turbo

This LoRA enhances your images with more punch, contrast, depth of field, and lighting. It's been trained on a mixture of photographic content with a focus on low-key, film noir, and fashion photography. It makes things pop without destroying image quality.

Suggested model strength: 0.3-0.8

Suggested samplers/schedulers:

• ddim / kl_optimal (or beta)

• dpm_2_ancestral / sgm_uniform (or ddim_uniform)

Version 2 updates!

This is a much more refined LoRA now. The convergence ended up being much better than v1 (this one is 0.336 compared to something like 0.71) so that means it's far more stable and consistent. I've adjusted the training and dataset so that it doesn't result in overly soft/blurry images anymore. This results in sharper images, but if you do want that soft focus effect you can still get it by adjusting your prompt to ask for it.

The usable range for the model strength is also much greater than before. You can go all the way up to 1.0 but you will find a sweet spot in the middle somewhere. This is pretty common for LoRAs for Z Image Turbo, best I can tell. I went through 5 or 6 iterations training this version and made sure that things like text weren't getting messed up. I wanted to make sure the LoRA wasn't fighting Z Image and taking away from its good qualities.

Bonus Tip! You can use both v1 and v2 LoRAs together for a stronger effect. Just be sure to lower the strength of the v1 LoRAs down to around 0.25 to 0.3 if you are doing so. You can try higher values, but you will run into quality degradation if the combined strength is too high between the two.

Quick comparison that does a great job illustrating the effect: