First: A quick public service announcement from a random person on the internet you don't know: smoking is bad for your health. Please don't smoke. <h3 id="heading-4">Update: Check out my <a target="_blank" rel="ugc" href="https://civarchive.com/models/89205/cigarette-smoker-woman">Cigarette Smoker Woman</a> for a more stable smoking lora.</h3> Now onto the description: I use this at 1.0 weight, though it will blend better with other loras at 0.9 or even 0.8. However, the lower the weight gets, the more it seems to do weird things. All images were tagged with <code>smoking</code> so it should be a reliable trigger. The <code>holding_cigarette</code> tag will try to push it more towards images with a cigarette in their hand instead of their mouth. Conversely <code>with a cigarette in their mouth</code> should push it more towards, you guessed it, a cigarette in their mouth. In both cases it will sometimes completely ignore that instruction for reasons I don't know. Adding <code>smoke</code> will make it more likely to generate a plume of smoke somewhere, and if you're lucky sometimes that smoke is even coming from their mouth or from the cigarette. Use lots of negative prompts to help weed out some of the common problems like mutated hands, blurry images, etc.Making a Lora for smoking is surprisingly difficult. This is as close as I could get after several attempts, and multiple days of GPU training time. It's definitely not close to perfect, but it can generate a usable picture about one in five times. I noticed someone else on civitai tried to make a model for smoking and ran into similar issues. I think the problem is that there are a huge variety of positions, grip styles, cigarette styles/brands, etc, that all qualify as "smoking". I kept adding images hoping that the model might eventually figure this all out, and ended up with over 130 images. The data set includes people of various genders, races, and ages, though admittedly it could still use more diversity. I had previously also included paintings and cgi art, but just a few of those skewed the whole Lora to producing only paintings, so the current data set only includes photos of real humans. The data set has cigarettes in the mouth, in the hand, holding it like a pen, holding it between fingers, holding it in their mouth with their hand, lighting the cigarette with a lighter, lighting it with a match, etc. I also used a handful of regularization images to refine what a cigarette even looked like in the first place (though it could definitely use more of them, that fixed some issues where a cigarette looks like a magic wand). After all that, it still has a hard time consistently producing a simple "woman smoking" image without a ton of positive and negative tags. I'm kind of exhausted with training this model and moving on to something else, but I may come back to it at some point. If someone else picks up the baton here, I might suggest trying to narrow down the definition of "smoking". Pick a specific grip style for example or if you want a cigarette in the subject's mouth then only train on images where the cigarette is in their mouth. The pros:<ul><li>It can usually make an object that looks like a cigarette.</li></ul><ul><li>Sometimes that cigarette is actually in their hands or mouth as you would expect.</li><li>It should (hopefully) work with a wide variety of people of different ages, genders, and races</li><li>It is often pretty good at making smoke, and that smoke is sometimes even near the cigarette</li></ul>The cons:<ul><li>Sometimes the cigarette is being held in an awkward position in the subject's hand.</li><li>Will often deform the hands and seems to make the model "forget" what hands should actually look like.</li><li>Will sometimes turn one or more fingers into a cigarette.</li><li>Will sometimes remove whatever object is being held in the other hand (like a cell phone) (reduce the weight to 0.9 or 0.8 to help fix that)</li><li>Will sometimes make multiple cigarettes (one in mouth, one in hand)</li><li>A cigarette in the mouth may be at a super weird angle, or not even align with the lips</li><li>The cigarette is often not lit or shows no ash</li><li>The smoke it makes is sometimes in the wrong place (smoke near their hand when the cigarette is in their mouth, or just in some random other place)</li></ul>Combining it with a TI that can fix hands might help with some of the hands mutating issues.

Smoking a cigarette

Trained on stock SD 1.5, but the example images you see here are produced with deliberate_v2. Data set Includes 134 realistic images of people smoking. The data set includes people of various genders, races, and ages. Also includes a handful of images (35) of cigarettes for regularization. The data set could still use more images, more diversity, and more regularization images.

Files with SHA256: b812a1859ea5994ec704385c18faec69c4ceed4308509c16d7ea4e3e67fc7364