AirbagDeploy V2_low_noise.safetensors

146 MB · SafeTensor

Requires membership

<airbag-low-wan>

H3 Ref Combo Model (Passenger+Driver) I went back on my previous "crashform" revelation and instead retrained with the airbag trigger and combined my Passenger and Driver dataset. So this one is easier to use and is very versatile. Be sure to designate "passenger airbag" vs "driver airbag" and note the deployment zone of each. "Steering wheel" works well enough for the driver airbag but you'll want to specify where on the dashboard the passenger airbag will deploy. H3 Crashform Versions (Fl2a-Ref-Passenger-Driver) Ok, so I had a revelation and retrained the driver model and a new passenger model. The issue with H3 is that it has a very solid idea of what an "airbag" is, so using "airbag" in the prompt/captions really muddled things up. I've gone ahead and replaced "airbag" with "crashform." YOU MUST USE "CRASHFORM" IN YOUR PROMPT INSTEAD OF "AIRBAG." DO NOT MENTION "AIRBAGS." If you do it will muddy the results and add tassels and other stuff to your airbag. Otherwise this driver one preforms WAY better than previous version. The passenger version I'm kinda shocked at how well it works because no model has been able to do it. Passenger one is more versatile, I think, but both respond well to shape and texture prompts, as well as deployment zones. For passenger I recommend "top of dashboard." H3 Version - Combo Combines both datasets though I kept them prompted for LTX/Wan. Will redo at some point. But it works as is, fairly straight forward to use. AirbagDeploy is the trigger but you don't really need it. H3 has a good idea of what an airbag is, so keep the strength I'd say no higher than .6. Higher gives you a more explosive deployment but fucks with animation. LTX 2.3 - Realistic Version This one was trained using actual crash test and dashcam footage. It creates a realistic crash and deployment, but it does suffer with the "structure" of the airbag. In that it can deploy pretty limp. I recommend pairing it with the "Puffy" version, but turning the weight down on the Puffy version to .1 to .15. In fact, the two LTX 2.3 versions pair well together and should be used in conjunction, shifting the weight balance as desired. LTX 2.3 Version Best version by far. Prompt for sound and dialog otherwise you'll get music. Works best at .75 strength. LTX2 Version This one is pretty straight forward. I don't quite have a handle on how to get quality outputs with LTX2 yet, but the lora does work. It wasn't trained on the sound of an airbag deployment (tried that and it was a nightmare) so you'll have to prompt it. Otherwise it, for some reason, adds in a horror soundtrack. The physical motion works great, and the airbags look good. It's versatile and FAR dynamic than the Wan 2.2 loras. Best results at .30-.60 strength. Higher and the sound goes to shit. EDIT: Version 2 works better at higher strengths. It also has a decent deployment sound without much work being done, but I would recommend prompting the sound. Version 2 Version 2 adds more dynamic motion to the character as well as a much faster and consistent bag inflation. It does also work well with post-inflation starting images for airbag interaction. It's pretty versatile and you can have an airbag come out of pretty much anywhere, but it still struggles with front views if a steering wheel is not clearly visible. Version 1 A lora to add airbag deployment to Wan 2.2. This is separate from my Airbag Hug model for 2.1 but can work in conjunction with it. I'll make a Airbag Hug 2.2 shortly. The starting image should, of course, be some one in a car. You can technically do a portrait and the video will shift to them in a car. I don't know how well it works deploying airbags in other settings.

SHA256

55E6CAA7B08B57ABB843D5A52216C02E68555D7B74D29151B24E06BC573636AC

ModelTrends.ai Model ID

#22755

Size

720x1280

Size

1360x720

Size

608x608

Size

1920x1080

Size

576x800