Stuck in Objects

V2.0 is here!

V2 is quite different from V1, and I attempted to massively simplify it for more flexible use.
The end result is the characters can get stuck in almost any object described (that I've tested) assuming the object is big enough. I also simplified the pose descriptions to use words already understood by Anima to help with posing.

ADVICE: if the pose isn't as you expected, just describe it in more detail, it usually works!

NOTE: The cube-bound pose from v1 does not work in v2 due it causing stability issues.

V2 Prompt Format:

[pose], stuck in [object] with [style] grommets

POSES: The main poses from v1 transfer over:

• wall-bound -> torso out

• spread-bound -> spread arms and legs

• behind-bound -> body out, arched back

• folded-bound -> folded

Other poses not trained for also work decently, such as on all fours or laying on back

OBJECT:

Just about any object works if it is big enough, but here are some examples:

• door, tree, urinal, wall, rectangular object, sphere, dresser, inside dresser, plaque, painting frame, vending machine

GROMMETS:

More flexible with different styles. mainly still gold, metal, rubber. Can be large

V1 Information:

A lora that helps with stuck in wall/object poses. an unofficial successor to the stuck in object lora by out_of_username_ideas (https://civarchive.com/models/1379252/stuck-in-object-wall-urinal-tentacles-tree-plaque-etc)

Can generate a number of "stuck in" objects, but works best with the trained dataset on machines, plaques, walls, and cubes.

Trained on base anima (https://civarchive.com/models/2458426/anima)

WARNING: I've noticed that the results from my own renders through ComfyUI are vastly different from my results using CIVITAI. I don't know how exactly they run Anima models so use at your own risk. You can download my images for the ComfyUI setup, or make sure to use these models:
unet: anima-base-1.0
clip: qwen_3_0.6b_base
vae: qwen_image_vae

Examples

The examples below are some minimal examples that help consistently get the trained poses

• woman folded-bound with visible head torso crotch ass and thighs, she is nude stuck in a silver shield shaped plaque on a wall with rubber grommets

• cube-bound, woman stuck in metal cube with metal grommets, visible head crotch feet and ass, nude, pussy and ass visible, her legs are sticking out of the cube over her head, her head is sticking out the top,

• spread-bound, woman stuck in machine with rubber grommets, visible head torso crotch,

• behind-bound, woman stuck in wall, visible head torso crotch,

Prompt Format

[pose] with visible [body parts] stuck in [object] with [style] grommets

Thanks to natural language, you can be pretty loose in the order of describing the concept, but it is works best when using the specific [pose], [body parts], and [object] you want.

Trained Poses:

• wall-bound - classic 'through wall' style image where the head and torso are sticking out

• spread-bound - spread arms and legs stuck in 2 walls on either side, usually in a machine.

• cube-bound - in a small object like a cube with body parts sticking out (a bit inconsistent)

• behind-bound - arms and legs stuck in the object with the rest of the body visible in a standing-like position.

• folded-bound - like behind-bound but more of a folded pose with the legs above the crotch

The poses give a good starting point but can sometimes blend into each other.

Trained Body Parts:

• head, torso, crotch, breasts, arms, legs, thighs, ass, hands, feet

• just describe them sequentially without punctuation, for example: visible head legs torso and crotch

• body parts help determine the pose, with parts like 'ass' and 'thighs' useful for getting the folded-bound pose instead of the behind-bound pose

Trained Objects:

• wall, plaque, machine, cube, metal plates (on wall)

Grommet Style:

Grommets are the rings that show up around body parts sticking out of the object.

• metal, rubber

Negative Prompting:

Using 'no [body part]' appears to be far more effective than having the body part in the negative prompt. So if you don't want to see the character's head, put 'no head' in the positive prompt.

Misc Considerations:

The model is probably a bit over trained, since i can get the strength down to 0.4 and it still works, but I haven't noticed any major issues with higher strength.

There is some noted training data bleed for: gags, dildo insertions (only for spread-bound pose), sweat and shake lines (especially if you start overcooking the image)

This model does tend to have a few style issues and over detailing when used with the anima turbo lora.

the lora was originally called "encased" (which is why I made that joke image of the girl in a suitcase) and the ComfyUI workflow embedded in my images will show that name instead of "stuck_in_objects", but they are the same.