NAI-Style Inflatables (PonyXL)版本v0.6.1 (PXL) (PXL) 综合资源合集综合资源合集

Version 1.0.0

After a lot of learning (and trial and error), this finally has reached a point where it has a decent grip on everything I originally aimed for. Hopefully others can get some mileage out of this too.

If you find any notable weaknesses or issues you'd like me to try and correct, feel free to reach out - but I can't promise anything.

Quickstart Tags:

Trigger word - plinflatable
Degree of realism - realistic, draw over, digital art
Material - matte, glossy, ultragloss
Perspective - low angle view, high angle view, three-quarters view, side view, front view, birsdeye view, oblique, tilted
Valves - valve on (location) - add "extra valves" to the negative prompt
Poses - on front, on back, leaning up, laying, standing, sitting, hand(s) on (location)
Water - floating, partially submerged, clear water

What's new:

Now fairly cooperative with portraits, not just full body images
Dramatically improved material tags
Shape tags - "slim," "toy proportions," and "thin arms" can help achieve (or discourage) a certain form factor

Known issues:

Does not play nicely with handles - tries to make them into valves
Lake backgrounds confuse the skyline with another waterline
Close-up backgrounds for realistic and draw overs are grainy

Other notable tags used in training:

Species - mizutsune, wickerbeast, dragon, moth, lycanroc, dialga, wolf, espeon, (alolan) vulpix, lion, kobold, dragonite, absol, lugia
Cloud cover - partly cloudy, overcast, storm clouds
Locales - beach, lake, ocean, river, mountains

Version 0.6.1

Many, many changes. Most for the better, some not so much.

What's the same:

Trigger word: plinflatable
Trained on: SFW NAI outputs of inflatables (but more this time)
Degree of realism/style: realistic, drawover

What's different:

More consistent generalization
Valve location tags - "valve on thigh" is now "valve on hip"
Perspective tags - now uses "X view" rather than "X angle" to hopefully improve consistency
Improved valve placement - more receptive to locations other than "hip"

Known issues:

Struggles with "zoroark" and "absol"
Still struggles with valve placement consistency, duplication, etc
Backgrounds are often muted and require some prompting
The usual culprits (limbs, etc)
Beach/lake backgrounds HEAVILY biased towards realism
Water can come out murky

Notable tags used in training:

Realism: realistic, photo, drawover, digital art

Perspective: high angle view, side view, three-quarters view, low angle view, birdseye view, oblique, tilted

Locales: lake, beach, ocean, river

Poses: laying, standing, on front, on side, on back, leaning up, sitting, hand(s) on cheek/chest/belly/etc

Props: leaning against rocks, lounge chair

Cloud Cover: partly cloudy, overcast, storm clouds

Material Finish: matte, glossy, ultragloss

Valves: no valve, valve on thigh/shoulder/tail/hand/belly

Original Description (v0.2)

A LoRA trained on SFW NAI outputs of inflatables (~75, x2 including mirrored versions). I'm new to this training thing, so don't expect much.

Generally has a lapped-seam-pooltoy look. Mostly trained on feral subjects.

Trigger word: plinflatable
Primarily uses tags. Should have decent understanding of the following:

Species: mizutsune, wickerbeast, dragon, moth, koi (somewhat)

Degree of Realism: realistic, drawover - biased towards realism

Should have vague understanding of the following:

Certain Pokemon: vulpix, lugia, lycanroc, dragonite, absol, a few others

Valve locations: Belly, thigh, shoulder, tail, no valves - it's fairly stubborn, though

Has frequent problems with the following:

The Classics: paws, feet, wings, etc

Valve Duplication: "extra valves"/"multiple valves" in negative helps somewhat

Handles: (the training set only had a few images with handles on thighs)

Color Bias: Red and purple are frequent culprits

I'm sure I'm missing something, but this will do for now.
Seems to work decently well with the couple other PXL inflatable LoRAs that I tried. May be worth experimenting with.