JUJUTUS KAISEN UNIFORM — Female JUJUTUS KAISEN Uniform Outfit LoRA (Jujutsu Kaisen) - PONY

JUJUTUS KAISEN UNIFORM — Female JUJUTUS KAISEN Uniform Outfit LoRA (Pony )
============================================================

Jujutsu Kaisen uniform, but make it a skirt.

This is an outfit LoRA for female characters wearing a JJK-style uniform: a cropped navy jacket with gold buttons, pleated skirt, black belt, and pantyhose. It's not a character LoRA — it's purely the outfit. Works with any character LoRA or on its own.

Created because Pony XL has a stubborn military uniform prior that's hard to override with prompts alone. This LoRA bakes the female JJK uniform concept directly into the model.


🏷️ TRIGGER WORD & COMPANION TAGS
============================================================

Trigger word:
jjkf_uniform

The trigger word alone is NOT enough on Pony. You MUST include these companion tags:

jjkf_uniform, cropped jacket, high collar, pleated skirt, black jacket, black pantyhose, black skirt, belt, high-waist skirt, small_head

Yes, the whole string. Pony needs all of these to fight its military uniform prior. The small_head tag helps prevent the head cropping issue from the neck-down training dataset.


⚙️ RECOMMENDED SETTINGS
============================================================

LoRA weight (model): 0.40 – 0.75
LoRA weight (CLIP): 1.0
Sampler: DPM++ 2M Karras or Euler A
Steps: 25 – 35
CFG Scale: 4

Tested checkpoints: WAI-ANI-PONYXL v7, Pony V6

IMPORTANT — LoRA strength is contextual:
- Solo use: can go up to 0.7 — strong and accurate
- Combined with character LoRAs: drop to 0.4–0.5 — this LoRA is very dominant, the outfit sticks even at low weight
- Above 0.75: outfit starts cropping like the training dataset (neck-down)
- Below 0.4: Pony's military uniform prior takes over; the uniform turns into a trench coat


📝 POSITIVE PROMPT TEMPLATE
============================================================

score_9, score_8_up, score_7_up, source_anime, rating_safe,
1girl, solo, jjkf_uniform, cropped jacket, high collar,
pleated skirt, black jacket, black pantyhose, black skirt,
belt, high-waist skirt, small_head,
[character tags: hair, eyes, expression],
[pose tags], white background, simple background, full body


❌ NEGATIVE PROMPT
============================================================

score_4, score_5, score_6, lowres, bad anatomy, bad hands,
text, error, cropped, worst quality, low quality, blurry,
source_furry, source_cartoon, 3d, realistic, ugly, male,
1boy, head out of frame, faceless, neck cut off,
pants, trousers, trench coat, long skirt, beige, cream

Additional notes: head out of frame, faceless, neck cut off — the training dataset was cropped neck-down, so the LoRA can learn to crop heads at high strength. These tags prevent that.


⚠️ KNOWN PROBLEMS
============================================================

Trigger word is not standalone. Pony has a strong prior toward "military uniform" and "trench coat". Without the full companion tag string, you'll get a long military jacket instead. Always copy the full companion tags listed above.

Head cropping at high strength. The training dataset is 60 images cropped neck-down (outfit focus). Above 0.75 weight, the LoRA becomes too dominant and the dataset crop behavior re-emerges — heads get cut off. Fix: keep weight at or below 0.75, and include head out of frame, faceless, neck cut off in the negative prompt.

Military uniform takeover at low strength. Below 0.4 weight, the LoRA signal is too weak and Pony's default military uniform wins. Fix: don't go below 0.4 for solo use. If combining with a character LoRA at low weight, reinforce the companion tags in your positive prompt.

Color drift to beige or cream. The uniform can sometimes shift to beige or cream tones instead of navy/black. Add beige, cream to the negative prompt. Training captions were filtered for consistent black/navy color tags.

CFG above 4 can cause cropping. At CFG 5 and above, the LoRA becomes more dominant and dataset crop behavior returns even with reinforced negatives. Recommendation: stick to CFG 4.

Character tags (hair, eyes) sometimes not followed. This is expected — the LoRA was trained on neck-down crops and knows nothing about faces or hair. Character appearance comes entirely from the base model + your prompt. Strengthen your character tags and add unwanted features (e.g., blonde hair) to the negative prompt.


📊 TRAINING DETAILS
============================================================

Base Model: ponyDiffusionV6XL (Pony SDXL)
Network: LoRA — Dim 32 / Alpha 16
Epochs: 15
Batch Size: 4
Optimizer: AdamW with Cosine LR scheduler
Learning Rate: 1e-4 (text encoder: 5e-5)
Resolution: 1024×1024 (bucket training)
Total Steps: 2,310
Final Loss: 0.0892
Dataset: 60 images (neck-down crops, female JJK uniform focus)
Captioning: WD14 Tagger — Danbooru-style tags

Why 15 epochs? Pony is extremely stubborn about military uniforms. It takes more epochs to plant "pleated skirt" and "cropped jacket" into the model. The text encoder LR of 5e-5 is critical — without it, the trigger word won't stick.

Why only 60 images? Outfit LoRAs don't need character or face variety. Variation comes from angles, poses, and lighting. 60 images is plenty for a single outfit.

Why Dim 32? Version 3 with Dim 16 failed — the military entanglement couldn't be removed. Dim 32 gives more capacity to fight Pony's prior and implant the JJK uniform cleanly.


✅ COMPATIBLE WITH
============================================================

Pony SDXL and its derivatives. Tested on WAI-ANI-PONYXL v7 and Pony V6. Not compatible with Illustrious/NoobAI — that needs a separate version (check my other models).


💡 TIPS
============================================================

Combine with character LoRAs — this is the main use case. This outfit LoRA is very strong. When pairing with a character LoRA, drop the outfit weight to 0.4–0.5. Run your character LoRA at 0.7–0.8 and the outfit at 0.4–0.5 — they work together without conflict.

Generate character datasets. Use this LoRA to generate images of female characters in the JJK uniform, then train them as character LoRAs. Useful for building a character series.

CFG 4 is the sweet spot. Above that, the LoRA oversaturates and dataset crop behavior returns. Below that, outfit detail gets softer.

Don't forget the Pony prefix. Always start your prompt with score_9, score_8_up, score_7_up, source_anime — this is mandatory for all Pony LoRAs.

If character tags are inconsistent (e.g., asking for black hair but getting blonde), that's the base model, not the LoRA. This outfit LoRA doesn't touch face or hair at all. Strengthen your character tags in the prompt.



🎨 COMMISSIONS & SUGGESTIONS
============================================================

Got a character you want to see in the JJK uniform? Need a specific outfit variation or a custom clothing LoRA? I take suggestions — and yes, I do commissions too.

If you want your own custom LoRA trained (character, style, outfit, pose — anything), hit me up. I'll work with you on dataset curation and training to get the result you want.

Drop a comment or DM if you're interested. Always happy to chat about new ideas.