For the most part, my LoRAs all get named 'Eldritch something-or-other' just cuz it's my username. There's little that's strange about most my models. But with this one, it's entirely appropriate. It's even part of my preferred prompting "Eldritch found photo". The model starts with some very old Midjourney images - v3 I believe. They're utter garbage in terms of realism of any sort. But they still have a compelling texture and atmosphere. Through a series of training those old MJ output on SDXL, Flux, and now Z-image, then generating with these models, recycling its output into new training data, we get a little closer to coherent imagery, while retaining the atmosphere and a good deal of the chaos that made those old AI output so compelling.Other terms you may want to add to your prompt: gritty damaged print (often leads to frames and borders), old tintype photo, grunge, and anything along those lines. I managed to unintentionally guide this model toward a super specific set of female features - to the extent that, for portraits of women, this might as well be a character LoRA. You can make her Asian, Mexican, Swedish, African ... the skin tone will change but the same features constantly show up. I'm perfectly fine with this.

Eldritch Chaos | for Z-Image

Eldritch_Chaos_Z_0.9.5.safetensors

v1.0

For the most part, my LoRAs all get named 'Eldritch something-or-other' just cuz it's my username. There's little that's strange about most my models. But with this one, it's entirely appropriate. It's even part of my preferred prompting "Eldritch found photo".The model starts with some very old Midjourney images - v3 I believe. They're utter garbage in terms of realism of any sort. But they still have a compelling texture and atmosphere. Through a series of training those old MJ output on SDXL, Flux, and now Z-image, then generating with these models, recycling its output into new training data, we get a little closer to coherent imagery, while retaining the atmosphere and a good deal of the chaos that made those old AI output so compelling.Other terms you may want to add to your prompt: gritty damaged print (often leads to frames and borders), old tintype photo, grunge, and anything along those lines.I managed to unintentionally guide this model toward a super specific set of female features - to the extent that, for portraits of women, this might as well be a character LoRA. You can make her Asian, Mexican, Swedish, African ... the skin tone will change but the same features constantly show up. I'm perfectly fine with this.

Files with SHA256: 26ff4724f658aed24b2d8ab5960021e059a2a2fc600d1d26f5dd3b9cac47c4f8