<h2 id="overview">Overview</h2>Generate the aesthetics of the everyday! This model was an experimental project designed to explore the impact of noise and compression artifacts on Z-Image. Unlike the more polished entries in my collection, this LoRA focuses on the technical "imperfections" that define modern mobile photography.<h2 id="compatibility-and-usage">Compatibility &amp; Usage</h2>While "Vacation" is in the name - reflecting the bulk of the training data - this model is effectively a style-shifter for realism. It injects a distinct "amateur-ish" feel into outputs, mimicking the look of smartphone captures rather than professional gear.Trigger Words: As always, 'trigger words' aren't real—just prompt normally! However, words like 'snapshot', 'smartphone photo', or 'vacation' appear frequently in the captions and can help steer the model toward that candid, casual aesthetic.Technical Specs: This is a Rank 16 LoRA saved with float32 precision.Strength: Optimised for Z-Image. If using Turbo, a strength of 1 is recommended as a starting point.All images generated at 1MP resolutions, then upscaled 1.5x with 0.5 refiner control percentage. Generated on Z-Image Turbo, 10 steps, Gradient Estimation Sampler, Beta 1.1 Scheduler. <h2 id="limitations">Limitations</h2>This is an experimental technical study on noise and compression. It was not designed for a specific subject matter, but rather to see how Z-Image handles the loss of fidelity associated with mobile sensors. It is not an NSFW-specific model, though it should remain flexible with the base model's capabilities.<h2 id="future">Future</h2>I am currently working toward a major Z-Image finetune called 'SoReal!' (or ZoReal!). My goal is to create the ultimate amateur-style finetune by utilizing:<ol><li>A custom-trained quality model.</li><li>A custom one-shot demographic model (ConvNext-XL) with 89% accuracy for skin tone, ethnicity, and body shape.</li><li>A finetuned wd-tagger-large-v3 on 50k hand-tagged images.</li><li>Gemini 3 Flash-generated captions utilising full EXIF and camera metadata.</li></ol>I want to deliver a high-generalisation, NSFW-ready model without the "baked-in" look of over-trained LoRAs. However, I am currently limited by computing and hardware costs. If you want to support the development of SoReal!, please consider following me on Patreon!## Dataset &amp; Training<ul><li>Dataset: 450 images, primarily sourced from personal smartphone captures and curated external examples of mobile photography.</li><li>Training: Trained at a Batch Size of 16.</li><li>Focus: Specifically targeted at analysing noise patterns and compression artefacts to move away from the "perfect" AI look.</li></ul><h2 id="licensing">Licensing</h2>If you'd like to release a merge of this model, please contact me.Made with &lt;3 By BitcrushedHeart

Files with SHA256: 0f125256e9d1644700408fd7bbe9b02830548e052f3c0106468508eba0efc5aa