This LoRA is trained on inflatable My Little Ponies and will generate other characters as inflatables.
A list of versions with links to them and the upvotes/downloads ratio (updated Jun 5):
v1.0 PDXL (1/5) (first version, made using Kohya_SS scripts. Uses "inflatable". Later versions were trained using OneTrainer and used different trigger words)
v2.1 dirlpinf PDXL (1/12)
v2.2 ddptponix SDXL (1/5)
v2.2ddptponix DreamShaper (SD 1.5, just uploaded)
Note: If the new censorship filter in CivitAI's generator claims you are trying to generate "minor content", DO NOT INSIST on generating with that prompt. It will restrict your account. In my case, I was told that removing the trigger word "dirlpinf" made the filter stop...
Use "dirlpinf" instead of "inflatable" (it means "DisplayDisplace IRL pooltoy inflatable"), although "inflatable" may also work (not specifically trained in this version).
Trained and tested with WAI-????-illustrious-SDXL v14.0
Does not work with "Illustrious XL".
"hongyitoy" (and "hongyi" in a few cases) doesn't do much, but may help with some characters.
Try "hongyitoy princess celestia dirlpinf, irl" vs "princess celestia, hongyitoy, dirlpinf, irl", especially if you are trying to generate humans (who will probably only become inflatable as anthro ponies, and that needs a little more strength).
A note on ponies: without LoRAs, WAI-????-illustrious-SDXL generates anime humans by default. "pony" makes them more pony, but sometimes it's small horses. "my little pony" can make them more MLP-shaped, but the style may conflict with inflatables. Also, "daybreaker, my little pony" will generate a black Daybreaker, but a much better one comes out of a rough description.
"horseplaytoys" and "dirlpinf float" may generate floats, but there are also "airtime" floats, which have printed wings and win often.
"fully inflated" is default and not used.
The dataset of "v2.0 dirlpinf WAI-???? ill" with a couple small fixes.
cfg over 3 causes too much contrast.
The higher the strength, the more valves appear, at least with a cfg that is too high.
Not tested much beyond finding those usable settings. Try v1.0 PDXL if you don't like it.
I haven't been using enough parentheses in my prompts! The LoRA may be much more flexible than I thought.
Here are some ideas of why it seems to be worse than v1.0:
It uses more masks
Some images were not brought back into the dataset
OneTrainer may have dropped some images.
I don't know how to use the Alpha setting in OneTrainer.
The trigger words do not use the "inflatable" knowledge PDXL already has.
The main trigger word is ddptponix
Captions improved over v2.1 dirlpinf PDXL, but seems to be worse. SDXL catches all the substrings of words like "hongyitoy", "horseplaytoys", "arniemkiidoorwallwindow" and tries to generate Chinese people, toys in boxes, horses and windows.
The main trigger word is ddptponix
see the version details for description
It was trained in multiple stages on a gradually improved dataset of over 200 photos of a few inflatable equines from MLP G4. Though it was intended to be SFW, some prompts involving humans or company names can result in ???? things that were not in the dataset.
Tested with Pony Diffusion V6 XL and ZavyFantasiaXL - PDXL 2.0.
The txt files contained "unpruned" tags describing the details at the same time, but there were no images to explain what the details are not. So many of the "inflatable" words don't really do anything, because that's all this LoRA knows.
full body is a must to prevent broken close-ups (some of the close-ups have portrait, eye focus, hoof focus, face focus), but might work anyway because of other tags
inflatable [pony name]: inflatable princess celestia
inflatable [pony type]: inflatable alicorn
inflatable pony, inflatable toy
for quadrupeds: feral inflatable
for floats: pool float (but you might get better floats without it)
irl, photo can improve quality
Poses: standing, lying down, fully inflated, inflated enough
Views: side view, rear view, front view
Valves (works sometimes, but apparently not trained enough): closed valve, leg valve, mane valve, neck valve, body valve, ???? valve, shoulder valve, valve pushed in
a version of Celestia has open smile
horseplaytoys and hongyitoys or hongyitoy were supposed to suggest shapes and poses, but horseplaytoys alone might make ????, and hongyi is very strong for humans in PonyXL.
The farther wing often turns out badly or is not visible. missing wing in the negative prompt might help.
Inflatables usually don't have crowns, so you can add crown to negative prompt.
Not tested:
nonpop is in almost every caption, twice, but it may worsen the quality.
peytral print and hoof shoes print were meant to say that the regalia are not 3D. Not tested.
inflatable seam, white seams
Sometimes, the results are much better with 35-40 steps than with 25 (DPM++ 2M Karras).
Most of the dataset is photos of
quadruped ponies made by Guangzhou Hongyi Toy Manufacturing Co., Ltd., mostly Princess Celestia, Princess Luna and Zecora posted by Hongyi or ArnieMkII. There is no Applejack in the dataset for the first version, but it can generate her.
Worship the Moon and Praise the Sun previously sold by Horseplay Inflatables (very few photos)
Feels like Heaven and Reach for the Moon pool floats sold by Horseplay Inflatables.
the Cadance float and a few other toys.
Most of the characters are by Hasbro and Lauren Faust.
The licensing terms of the base model restrict the licensing terms of this LoRA. See https://civitai.com/models/257749/pony-diffusion-v6-xl
for WAI-NSFW-illustrious-SDXL v14.0
Not for Illustrious-XL 1.0
trigger word (optional?): dirlpinf
LR = 0.0003, 4006 steps. 3333 is 0.832 of that (strength 0.8).
Base model's recommended cfg: 5-7.
Filename: infpony-20250518-wainsfwill-2.safetensors
1. Hak untuk model yang diposting ulang adalah milik pembuat aslinya.
2. Pencipta asli yang ingin mengklaim model harap hubungi staf SeaArt AI melalui saluran resmi. Klik untuk mengklaim
