Z-Image model:recaptioned using Ollama/Qwen-vi, somewhat weaker than previous models. Use natural language captioning, "vintage mustang II" works well, as it is the term used in the captions, mentioning years between 1974 and 1978 might help keep some of the styling consistent. terms like "mustang logo" and "Ford logo" help make the badging clearer.Don't really need the htk-mstng trigger work in the Z-Image model.An Improvement on the first model.Added more images to dataset.Better captions:<ul><li>captioning included tags for year, bodystyle (hatchback, notchback, ghia)</li><li>also captioned headlights, grill, tail-lights, as well as general camera angle (view from front, view from behind, view from left, view from right)</li></ul>use htk-mstng2 to trigger.Usage:<ul><li>Use notchback/hatchback to not get weird SUV/crossover.</li><li>Use at normal 1.0 strength to get normal images.</li><li><s>Reduce strength to use to get that 70's mustang / Ford look to other things. or to try to include other subjects. Use part prompts like grill, headlights,</s></li><li>use with style/material LoRas to get odd variants. (the reason I created/ uploaded this)</li></ul>Experiment: added interior shows to the dataset. interior shots use the tag htk-mstg2intnot great, but makes an attempt to give interior. will likely explore this concept further.Overall I am happy with this one. Is shows some versatility, even attempts to be in a image with other vehicles, and works with other LoRas like hovering car, which is one of the reasons I made this.

Mustang Gen 2 (74-78)

Re-captioned the source images with qwen-vl. Then corrected the captions before training.Stopped training when it started to affect the prompt "vintage mustang" (as opposed to "vintage Mustang II". it never affected "car". Does interior shots better, but ZIT doesn't seem to understand a 2 door doesn't have doors for the back seats.

Files with SHA256: df331217ff2e2737d1904a7ead866e0a3ebb17b0efc16e126e1a368134817862