I tried to make one final version of this concept and yet I still think something is missing. I will left dataset source open so you can make your own versions.

Main prompt: sleepsack [add color sleepsack as additional prompt for specific color]

Optional prompt:

• rubber sleepsack [makes material more glossy]

• inflated sleepsack [makes sleepsack more puffy and large, color might affect inflation]

• belts [or color belts for specific color]

• buckles [or padlocks]

• zipper

  • latex hood, faceless, face hidden [should make faceless version of hood]

  • latex hood, eyes visible, mouth visible [the opposite of regular hood]

    • opened latex hood [face revealed in general]

      • mouth gag [face revealed in general except mouth]

  • gas mask [should make gas mask but do not expect too much the result]

  • plug gag, ball gag, gag, gagged

I would have recommend to use additional lora for hood/gasmask or inpaint for smooth experience. Dataset based mostly on sleepsack, so without supportive lora, result may be not good enough.

• chains, chained, straps [add chains to belts or instead of belts usually goes with hanging prompts]

• hanging out, hanging from ceiling, above floor [if you want them hang above ground]

• upside-down, upside down [for hanging upside down]

Negative (optionally): nothing important, maybe: english text, speech bubble [but chance is very low], or arms [if they keep appearing] or anything else you don't need

Do not forget to use Adetailer and Hires fix!

Lowering denoising strength may also generate better results.

Parameters may vary. Experiment by yourself.

The results of your generations are dependent on the additional LoRAs, weights, and models you use, so it may not work or come out as consistent as my sample images.