Aethon Qwen Furry Preview
Details
Download Files (1)
About this version
Model description
Something I tackled to learn more about LoRAs and Qwen's architecture. It can do ref sheets, stickers, fursuits, SFW and NSFW, and so on.
It needs long complex prompts. I didn't want to break what's good with Qwen like handling complex prompts and text by overwhelming it with too much fandom terminology. At least not at first.
Let me know if you like it or find any LoRAs that pair well with it =3.
Usage
Resolutions
| Aspect Ratio | Resolution |
| ------------ | ---------- |
| **16:9** | 1664×928 |
| **3:2** | 1584×1056 |
| **4:3** | 1472×1104 |
| **1:1** | 1328×1328 |
| **4:3** | 1104×1472 |
| **3:2** | 1056×1584 |
| **16:9** | 928×1664 |Settings
| Setting | Range | Starting Point |
| ------------------ | --------- | -------------- |
| LoRA Strength | 0.8 – 2.0 | 1.0 |
| Steps | 20-60 | 30 |
| Cfg | 2 – 7.5 | 3.5 |
| Sampler | ??? | Euler |
| Scheduler | ??? | Simple |
| Lightning LoRA | Untested | No |Prompts
It's safe to go extremely detailed. I'm not sure how to describe it but these are prompts I usesd when testing it out.
Reference Sheeet
A three-column reference sheet, front view (left), paw pad (pink on black) detail (center), back view (right); clean mint background, thin gridlines, rounded empty label boxes; anthro cat “Domino”; white fur, black eye mask connected above nose, black ears (pink inner), green eyes, hands to elbows, white digitigrade legs with thick thighs that turns black below the knees starting on the outside of the leg then covering his entire calf and foot.
The character's proportions are those of a 24yo male and have digitigrade legs with thick digitgrade leg white thighs. He has hand and foot paws have 4 fingers and black tips. striped tail, small shoulder spot, thigh marking, aa maraking at the top right of the front of his chest that's oblong and irregular; flat colors, by a professional artist.
In the front view he's waving at the camera with 1 hand and his mouth is open in a disarming, friendly smile, in the back view, he's making fists.
The paw detail occupies most of the space in the center with negative space surrounding it. It shows the bottom of a black cat paw with 4 pink pads, "Domino" as the headline. Below the paw detail, the lines "Domestic Cat" "He/Him/His | Level 24" "A huge dork and your best friend..."
Stickers
This didn't work as well as I'd hoped. I think it needs more stickers in the dataset and more specific terms like "hugging a generic blue YCH woof."
a gay male furry anthro fox is reclining in bed, propped up by pastel color pillows seductively. He's stroking erect, canine penis, below which are his testicles, perineum, and anus. The fox is 2 tone mottledd maroon on his back, sides, top of his head, a cream color belly. NSFW furry art. The penis is dripping white precum. One leg lifted up, maroon paw pads. Vector art, telegram sticker with a white border on a simple light purple background. heterochromia with one blue eye and one yellow eye. eyelids partly closed. he has 2 sets of testicles. one in the bottom 14 of his erect penis. his legs, thighs, and butt are maroon
A comic bubble above him saying "did you forget something?" his paw pads are glossy
Fursuit
A gay man is dressed as a blue and black anthropomorphic tiger with digitigrade legs waiting in line to get on an elevator in the lobby of a Westin hotel. His fursuit is bright and features an orange border around the white belly. He's attending a furry con and there are 3 people in costumes in the scene as well as a few in normal casual clothes.
Female Horse Sketch
In another, I tried a species and a gender that's absent from the dataset (because reasons) and got good results with just below, but had to turn up the LoRA strength past 1.0.
a full length digital sketch of a furry anthropomorphic horse woman in a bridal gown in a beautiful, peaceful meadow.
Commercial Use
No restrictions, but please don't use this to compete with Furry artists directly. Working with others on a sona is an important part of the culture. Other anthropomorphic work or commercial clients are fine.
Tech Notes
Trained at bfloat16 with float32 text encoder and VAE, 1328px resolution batch 4, gradient 2.. About 500 steps under prodigy with cosine scheduler, 300 steps of AdamW with EMA to stabilize, and another 100-200 at a low LR for annealing. About 17hrs on an H100 with a 600 image dataset and 120 validation images. Rank 64, alpha 64. I used OneTrainer's cloud tool with GPU from vast.ai
Captioning by JoyCaption with category specific prompts. The dataset started at 3,800, but was pruned by clustering using Nomic 7B embeddings and FiftyOne.










