Flux Klein Version 2

A passenger version will be coming out shortly. Same rules apply.

This one was trained with prodigy. Very versatile and with a much better texture. It's also way more finicky so specificity is the key. It has a habit of swallowing body parts and items, so prompt where you want each one to go. It also helps to position the airbag in the scene with directions (i.e. if your steering wheel is on the left of the frame and the woman is on the right, mention that). Prompt which items are interacting with the airbag too. IF the airbag is naturally too big the lora is very responsive to size directions.

Flux Klein 4B/9b

With 4B Detail is king here, so describe exactly what you want, including arm placement, body placement, and airbag texture.

9B is far more forgiving, but generally try to be detailed.

Version 2 - 2511

This one was a test of a new captioning method I used with this workflow.

This version is by far the most dynamic, with the airbag pretty much doing whatever you want it to do. Body interaction is more versatile, and it doesn't lock the character into a position. I don't have many tips other than to use an LLM to direct the prompts if you can.

Otherwise, have fun!

Version 1 - 2511

Works pretty much the same, except with much better versatility that 2511 gives you.

Several poses are baked in. "Make X hug the airbag" makes her hug it. "Make x push the airbag away" will make her push the airbag away, though it helps if you specifiy that there should be space between the character and their body. "Make X face plant into the airbag" will make her bury her face in it. "Make X use their body to press the airbag against the steering wheel" will have the character lean forward.

But like I said, very flexible. You can use it however you wish; adding an airbag to a scene or using a character image to put a character in an airbag scene.

Version 4 - 2509

I didn't like the texture of Version 3 so I trained Version 4 using the dynamics images from Version 3 and the higher res/better textured images from Version 2. I also trained it to 5k steps (the 4250 iteration is the best as far as I can see). It's a bit more ridged and VERY strong but if you turn down the strength it becomes far more flexible. At Strength 1.0 you can pretty much throw an image of any character/person in there and turn it into an airbag image.

Version 3 - 2509

Used a data set pulled from airbag videos I made. That leads to a softer and more realistic airbag and can be used for more dynamic poses.

Key phrases are as follows:

• "Make airbag deploy from steering wheel" - A round airbag will come out of the steering wheel, defaults to off-white. Add color indicator if needed.

• "Make girl faceplant into the airbag" - Makes the girl lean forward and press her body and face into the airbag.

• "Make girl push the airbag down" - Hand interaction. Works best with head on clips.

• "Make the airbag pin the girl to her chair" - Pushes girl back, better face and body coverage

2509 Version

This one was trained on the 2509 model using the same dataset from version 2. Works way better, far more versatile, and does not requiring drawing an airbag in there. Get creative with the prompts, it'll put an airbag where every you want. You'll want to be detailed when it comes to body interaction.

It is biased towards having a steering wheel in the shot, so you may need to get creative if you want to do passenger shots.

Version 2

New version was trained with 3k steps instead of 1.5k, though using the same data set. Works better. Works decently well with just plain image, but works very well if you draw a crude airbag in first either.

Version 1

Pretty versatile lora that allows you to insert an airbag into a scene. Works either inserting it into an image directly, or using a character sheet to generate a new image. The examples (of course) are from AI images I generated it, but as I said, it's versatile.