🎶 I feel the Rush.. addicted to your touch 🎶... of room deodorizing freshness!

FLUX Version 2: Based on a much better dataset in quality and quantity with a different captioning strategy - This version is much more flexible and greatly reduces any body horror with complex compositions/scenes than version 1.5, so you won't need to do much, if any, img2img corrections. If you're going to make a nude generation, a peen LoRA or peen-friendly checkpoint is highly recommended. Images in the model example gallery are txt2img examples.

Z-Image Turbo Version: Trained with an upgraded dataset with only HQ images, approx 356 total, with a simplified captioning strategy than what I used for Flux. With Z-Image as the base, the results are very good with prompt adherence and not much in body horror. All the images posted in the model gallery are one-shot txt2img generations without follow-up img2img inpaint or upscaling corrections.

FLUX Version 1.5: I trained using the same training object-in-context approach with my other recent FLUX LoRAs, but this time testing it out with a smaller dataset to figure a slightly different workflow. LoRA size still an issue but I'm testing a few things out so maybe future versions I'll have it figured out.

Anyway, back to this LoRA: The results were a fairly stable LoRA that plays well with character based LoRAs but results may vary depending on how the character based one was trained; I noticed very few of my character LoRAs did not want to play along w/ this one at all while most other ones were perfectly fine using it. I've included a few in the gallery as examples (btw - the gallery shows 3 for the cfgscale, but I actually used 3.5. For some reason Civi's image edit UI doesn't accept numbers with decimals).

A side 'benefit' is, since about a 3rd of the dataset included men clapping their cheeks, it can sorta generate full on sex scenes but don't expect any one-shots if you want to try to do that. If you do try that out, then a workaround is just correct any body horror from the first shot with an img2img inpaint process using the same or similar image generation parameters. I might work on that issue with later versions, if I grow the dataset. A second alternative workaround is that you can just generate the sex scene without it and then come in afterwards with an inpaint pass using it.

For those of y'all that still have or remember those days of VHS VCR players needing head cleaning to make those films pop.. since the 70s, 80s, 90s and thru today! Sometimes seen used in those trouble room areas of a busy house, like the bath!

--Pony v6--

trigger: poppers, smelling, sniffing, poppers, holding poppers,

label color can be specified and some brand labeling (real and not) exists with Rush being the most consistent.

--Pony v5--

trigger: poppers, smellingpoppers, sniffingpoppers, holding poppers

sniffing should trigger a finger against a nostril .. most of the time at least.

--SDXL v1--

trigger: junglejuice

*IYKYK