This LoRA adds and enhances digital glitch effects to your images, but you'll need to add some keywords to your prompt. While it has been trained with a trigger word, the effects are better controlled when adding more phrases to your prompt, ie.; digitally glitched, glitch effect, pixel sorting, data streams, digital noise, chromatic aberration, and so on. The trigger word is gl1tch and it can help a little bit too, but is completely optional.Suggested model strength: 0.9Workable model strength range: 0.1 - 1.2Suggested sampler/scheduler: ddim / beta (or euler / normal or beta)Z-Image Turbo will actually add some glitch effects to images without a LoRA, but it will be very minor. The Z-Image model really fights against distortion and noise so if you want very pronounced glitch effects (and consistently) you'll need a LoRA. This was not an easy LoRA to train. I'm sharing it now after it's 5th generation. It ended up with a training loss of 0.3032, so right before it ran into the danger zone of being over fit. The save point before the one I'm including was 0.348 and the difference between the two is pretty tiny, but technically 0.3032 is better.You can technically use this LoRA with the model strength set all the way up to 1.2, but it will begin to take away from subject details. Skin tones won't be fried and there won't be weird artifacts...at least not often and not major ones, remember the style is adding distortion after all, so imperfections are a bit hidden. You will normally want to use this at model strength around 0.9. 1.0 is very often perfectly fine, but I always like reducing the strength of Z-Image LoRAs. You can certainly dial in the glitch strength with this one though, so try different values. At 0.1 you won't see much effect at all, but it is distinguishable from no LoRA at all.For my suggested sampler/scheduler, I often prefer ddim with kl_optimal but that can often be too strong for this LoRA. Using ddim with beta scheduler is often better. This is very similar to euler sampler with normal scheduler though. I would probably avoid euler ancestral with this LoRA, but it can also work.

Files with SHA256: 8a38a29b7c2703dac19a245406788604d29e3c1b6ceab961a8e5cb387356866b