Since WAN2.2 is already a great model but only does have a rather basic understanding of the act of smoking cigarettes, I decided to train a LoRA for it.
The LoRA is trained on women only. It may work on men too, when using only the high noise LoRA. Although I can't guarantee it. Cigars will also work when already visible in the starting image.
No support for pipes, vapes, etc. - sorry.
NEW:
Version V4 now supports many more styles and tricks, like snapinhale, frenchinhale, openmouth-inhale, nose-exhale, smoke rings and lighting up a cigarette. Just use a triggerword (see details of the LoRA) and/or write a short text prompt.
Example prompt (without triggerword):
A young woman holds a cigarette, brings it to her lips, and takes a very strong, deep drag. First, she spreads her fingers and then removes her hand from the cigarette entirely; it remains held between her pursed lips without support as she draws on it so vigorously. Her cheeks hollow out deeply and her face appears very slender as she takes the long drag. The glowing tip of the cigarette flares brightly as she inhales hard. She elegantly removes the cigarette from her mouth and opens her lips wide, allowing a billowing cloud of smoke to flow slowly out. She inhales the smoke by quickly drawing it back into her lungs. She smiles slightly, then—with pursed lips—slowly blows a long, well-formed stream of thick smoke into the air.
Example prompt (with triggerword):
Nose3xhale, A young woman is holding a cigarette in her hand. She puts a cigarette between her lips. As she takes a very deep drag from that cigarette, her cheeks start caving inward and the tip of the cigarette glows. She takes the cigarette out of her mouth, inhales the smoke in a very rapid action and shuts her mouth. After holding the smoke within her chest for a moment, her lips remain shut and she begins to exhale — not through her mouth, but exclusively through her nose. Two fuzzy streams of faint vapor emerge slowly from both her nostrils, forming intense, focused jets that contrast with the environment. The smoke billows downwards in two directed very faint plumes as she finally laughs.
Description
Since WAN2.2 is already a great model but only does have a rather basic understanding of the act of smoking cigarettes, I decided to train a LoRA for it. The LoRA is trained on women only. Don't know if it works on men, too.
Strength high-noise: around 1.0
Example prompt:
A young woman is holding a cigarette in her hand, She puts the cigarette between her lips. Then she takes a very hard and deep drag from that cigarette. Her fingers first spread and then she removes her hand completely from the cigarette. Just her lips are holding the cigarette while she sucks so hard on the cigarette. Her cheeks are deeply caved inwards and her face becomes very slim and long while dragging so hard. The tip of the cigarette starts glowing yellow as she drags hard. Then she takes the cigarette out of her mouth, opens it and a wafting cloud of smoke becomes visible in front of her open mouth. Then she inhales the smoke by quickly sucking it into her lungs. Then she blows out a huge and long stream of thick smoke from her mouth into the air and then laughs.
FAQ
Comments (14)
Please give an example of prompt. Thank you.
Sure. Edited the description and added an example prompt.
Don't use the low noise model. Say "They are smoking." or "They are smoking with style." The low noise model ruins the output in my experience.
use the prompt I provided. and be specific on details. high lora strength around 1 and low at 0.5
@Duke_Silver IMHO the low noise model is definitely needed for details. Although only at strength around 0.5
@helo7 You provided that prompt AFTER I made my comment... At least, the prompt wasn't there until after I made my comment. To be fair, the site was having yet another MASSIVE issue at the time. Also, don't use such detailed descriptions of the action you are trying to capture. You are working against yourself. Use a trigger phrase and then caption all of the things you want to be able to change, not the motions you are trying to capture in your model.
@helo7 The low noise model in not needed and makes every male character look feminine. Outputs are better without it.
@Duke_Silver Yes, I'm fighting almost the whole weekend with that **** site. Not the best time to upload a lora for the first time here I guess. Maybe you're right not to use the low noise model on men. But like I wrote in the description, it was intended to be used on women only. So fair enough...
@helo7 Please don't take my comment as a criticism. The model is good, for sure. Nice work. It is just more flexible without the low noise model. Wan is an odd duck sometimes.
You're missing files in the low Version, and have two files in the high version... can i assume that the bigger file is the high and the smaller is the low?
vice versa...low is the bigger one. had some problems to upload both. sorry
unable to download the Loras "Error getting file"
seems civitai had some problems. download is working now.
The key to this one is to not use the low noise model. The high noise model does everything you need. Also, smoking is not sexy. Have fun!