Dark Atmospheric has been one of my all-time favorite LoRAs over the years. I usually pair it when using my own LoRAs. Excited to see it for Z Image. I did some comparisons here after seeing a conversation on the model page for this LoRA. I think there's a few things to note.

1. I'd keep the LoRA strength for Dark Atmospheric at <= 0.5 and this is fairly typical for most Z Image LoRAs. You start to fight the model and things result in lower quality. The LoRAs that end up better loss can be dialed up, but I don't often go beyond 0.8 for even the latest one I trained which has excellent numbers here. You can see the image in this post that had the Dark Atmospheric LoRA with a 1.0 model strength for an example, not difficult to identify. I also have in this post images generated with my LoRA at 0.8 and 1.0 for comparison.

2. Experiment with the sampler and scheduler. I've only found a few combinations that work with Z Image so far. euler, euler_ancestral, dpm_2, dpm_2_ancestral and some of the other dpm samplers but not all of them. I use either normal scheduler or ddim_uniform or sgm_uniform. These settings will often make a big difference.

All of the images in this post use the same prompt. I didn't remove the trigger word for my LoRA from the others, but had the ComfyUI node bypassed. The trigger word doesn't throw anything off, I checked. I also tried images with both LoRAs applied. This can work, but you have to lower the strength of both LoRAs quite a bit and the resulting image isn't really much better in terms of "darkness" so there's not much point.

Which is all to say, are LoRAs really necessary with Z Image? My answer is that it depends. You have to experiment and see if it fits what you're going after. I would say that in general Z Image is one of the best models for realistic photographic images I've seen yet. A lot of the LoRAs that were trained over the years have been to increase the realism for models and with Z Image, I don't think you'll find the need to do so as often. However, the prompting style has changed a bit from Flux. Take a look at the prompt I have for the images in this post.

You can easily send conflicting signals the model with your prompt. A word like "vibrant" will boost brightness across the entire image to make colors pop. "Ethereal" can often add a glowing, misty, or heavenly lighting effect that washes out shadows. "Contrast" can often result in things being clearly lit and thus brighter, not always, but sometimes. Words like "soft light" can be useful when you want to avoid crushing the blacks and having things pitch black, providing some detail to the dark areas, but it can also result in things becoming brighter than you wanted.

You can see what "ethereal atmosphere" did to the image here: https://civitai.com/images/112702685 ...which is a rather cool effect, I wasn't expecting that, but it wasn't what I was going for.

Try words like "pitch-black" and "enveloped in deep shadow" and "strong spotlight" if you want pockets of light rather than the entire scene.