My attempt to fix how anima handles fishnets.

Detaided dataset of 54 images. The initial prompt is retrieved using my grabber. In addition, fine details are inferred with Huihui Qwen3.5 27B Abliterated GGUF Q4_ .
Prompt:

See the classification of the fishnets:

For fishnets, the biggest mistake is to tag only “fishnet stockings” and ignore the attributes that actually control appearance.

A practical categorization:

1. Garment type

This tells the model what object it is.

fishnet stockings

fishnet pantyhose

fishnet tights

fishnet thighhighs

fishnet kneehighs

fishnet bodystocking

footless fishnet tights

open-foot fishnets

This matters because “stockings” vs “pantyhose” vs “thighhighs” changes the silhouette a lot.

2. Coverage / height

This is one of the attributes you mentioned.

knee-high

over-the-knee

thighhigh

hip-high

waist-high

high-waisted

low-rise

reaches lower abdomen

ends at hips

underbust bodystocking

For anime data, I would standardize this more cleanly as:

kneehigh_fishnets

thighhigh_fishnets

pantyhose_fishnets

lowrise_fishnet_pantyhose

highwaist_fishnet_pantyhose

Because “go to lower abdomen” is a bit vague for training tags.

3. Mesh size / cell density

This is absolutely core.

micro fishnet

fine mesh fishnet

medium mesh fishnet

wide mesh fishnet

large diamond fishnet

extra wide mesh fishnet

I would use about 4 bins only, otherwise annotation becomes inconsistent:

micro_mesh

fine_mesh

medium_mesh

wide_mesh

You can also tag shape if needed:

diamond mesh

square mesh

hex mesh

But diamond is so dominant that shape may only be worth tagging when unusual.

4. Color

Also core.

black fishnets

white fishnets

nude fishnets

brown fishnets

red fishnets

pink fishnets

blue fishnets

green fishnets

patterned fishnets with colored threads

For training, tag actual visible color, not assumed garment name.

5. Opacity / visibility of the threads

This is often missed.

sheer fishnet

opaque-thread fishnet

glossy fishnet

matte fishnet

translucent

bold contrast netting

Why it matters: two black fishnets can look very different if the threads are thin and subtle versus thick and high-contrast.

6. Thread thickness

Very useful and often confused with mesh size.

thin thread fishnet

medium thread fishnet

thick thread fishnet

A wide mesh with thin threads looks very different from wide mesh with thick threads.

7. Edge construction / top finish

Important for where and how they end.

plain top band

lace top

elastic band

scalloped top

reinforced waistband

visible waistband

hidden under skirt

garter-compatible top

For thighhigh fishnets especially:

lace-top thighhighs

plain-band thighhighs

garter fishnets

held by garter belt

8. Support method

Also important if they stop on the thigh.

suspenders

garter straps

garter belt

stay-up stockings

unsupported stockings

slipping stockings

This changes the visual storytelling and helps separate “thighhighs” from “gartered stockings.”

9. Condition / wear state

Very important for expressive variation.

intact fishnets

torn fishnets

ripped knee

laddered fishnets

stretched fishnets

sagging fishnets

uneven fishnets

worn fishnets

frayed edge

This is worth explicit tags if your data contains it.

10. Fit / tension

Often overlooked, but visually important.

tight fit

loose fit

stretched over thighs

slack around ankles

wrinkled at knees

bunched at ankles

compressing skin

Fishnets look different when highly stretched, because the diamonds elongate.

11. Body-area emphasis

Useful if your dataset is fashion-focused.

visible waistband

emphasis on hips

emphasis on thighs

emphasis on knees

emphasis on calves

emphasis on ankles

exposed toes

reinforced toe

open toe

12. Layering

Very useful.

fishnets over bare skin

fishnets over colored tights

fishnets under shorts

fishnets under skirt

fishnets with underwear visible

layered with stockings

layered with socks

layered with leg warmers

This strongly affects appearance and can confuse training if untagged.

I am going to need you to infer these attributes from the images I send you. Do the fishnets only. The output shall contain one line of tags only.

May we start?