Sign In

Sticks the Badger (Sonic Boom) — Krea 2

2

Download

1 variant available

fp32 SafeTensor

sticks_krea2_v1.safetensors

Full precision, largest file • 357.05 MB

Verified:

Type
LoRA
Stats

37

2

2

Reviews
Published

Sep 1, 2026

Base Model

Krea 2

Training
Steps: 1,440
Epochs: 30
Usage Tips
Strength: 1
Hash
AutoV2
B6813D75D4
Trigger Words
stkbdgr
default creator card background decoration
Followers - 8

8

Likes - 14

14

Downloads - 97

97

03_village_medium.png

Trigger word: stkbdgr

Sticks the Badger from Sonic Boom, trained on Krea 2. Outfit, framing, lighting and style stay promptable, and it stacks cleanly with realism LoRAs — the "captured IRL" photos in the gallery are the same LoRA with a realism stack chained after it.

Use stkbdgr, never the word "sticks". "Sticks" is an ordinary English plural and pulls the base model's own concept instead of the character.

Read this first

This targets attention layers only, which is what lets it compose with other LoRAs but it also means the render comes out as Krea 2's interpretation of her design, not a faithful flat-shaded Sonic Boom cel look. Sonic Boom's shading sits a long way from Krea 2's prior. You get her identity, her silhouette and her palette, rendered a bit richer and softer than the show. If you want a screen-accurate cartoon frame, this LoRA might not satisfy you fully.

Settings

Exactly what the gallery images were made with:

Model Krea 2 Turbo

LoRA weight 1.0 (0.8 when layering other LoRAs)

Steps 8

CFG / guidance 1.0 (CFG off — Turbo takes no negative prompt)

Sampler euler

Scheduler simple

Denoise 1.0

How to prompt it

Captions all followed one shape, so prompts built the same way land closest:

stkbdgr, an anthropomorphic badger with tan and brown fur, wild dark brown hair and large blue eyes, wearing [outfit], [pose / action], [lighting], [background], [framing], tv still

stkbdgr must lead. The phrase an anthropomorphic badger is in every caption, so including it helps.

Framing words it knows, with how many of the 48 captions used each: medium shot (22) · close-up (15) · wide shot (12) · full body (10) · head and shoulders (7)

tv still is the style tail. It ends every training caption, so it acts as a style switch, keep it for the show look, drop it when stacking a realism LoRA (leaving it in fights the realism and drags the image back toward CG).

Outfits it knows explicitly

  • wearing a tan cropped top and a grey-brown wrap skirt with gold arm bangles that's her default (34 images)

  • wearing a black sleeveless dress with white trim — the formal dress (8 images)

Example prompts

1 — portrait (the cover image)

stkbdgr, an anthropomorphic badger with tan and brown fur, wild dark brown hair and large blue eyes, wearing a tan cropped top and a grey-brown wrap skirt with gold arm bangles, head and shoulders portrait, three-quarter view, confident smirk, soft warm daylight, plain dark grey background, close-up, tv still

2 — character sheet, white background

stkbdgr, an anthropomorphic badger with tan and brown fur, wild dark brown hair and large blue eyes, wearing a tan cropped top and a grey-brown wrap skirt with gold arm bangles, standing upright arms at sides, neutral expression, plain white background, full body, clean studio render

3 — an outfit she has never worn

stkbdgr, an anthropomorphic badger with tan and brown fur, wild dark brown hair and large blue eyes, wearing polished steel plate armour, standing in a castle courtyard holding a wooden staff, overcast daylight, medium shot, tv still

4 — the second canon outfit

stkbdgr, an anthropomorphic badger with tan and brown fur, wild dark brown hair and large blue eyes, wearing a black sleeveless dress with white trim, seated at a candlelit dinner table looking at the camera, warm interior light, medium shot, tv still

5 — "captured IRL" candid photo (the showcase look)

The character stays stylised; the photograph carries the realism. Asking for a "photorealistic badger" instead collapses her into an actual animal.

stkbdgr, bodycam footage, slightly blurry and overexposed, a badger woman with anthropomorphic features in the real world, tan and dark brown fur, wild dark brown hair and large blue eyes, wearing a purple shirt over a white tee and cut-off denim shorts, pushing a shopping cart across a supermarket parking lot at golden hour, laughing and pointing off-camera, warm low sun flaring into the lens, rows of parked cars and a red store facade behind, cowboy shot from across the asphalt, early 2010s phone photo quality, heavy digital noise

LoRA chain (order matters, all model-only), 12 steps for this one:

Krea2-realism-V2 0.30

lenovo_ultrareal_krea2 1.00

nicegirls_krea2 0.40

sticks_krea2_v1 0.80

Always name an outfit in these prompts. Without it she often generates naked or with broken clothes.

Known limitations

  • Full-body proportions run short. At full-figure framing she comes out stockier than the show — the head reads too large and the legs too short. The dataset is dominated by medium and close shots, so the model learned her head far better than her legs. Two things help: generate full-body at around 896x1024 rather than a tall narrow aspect, and say full body explicitly. Medium and close framing are unaffected — that is where this LoRA is strongest.

  • Jewellery is not faithful. Her gold ring pendant, shell necklace and gold arm cuff come out as generic grey tribal pendants and vary per seed. Attention-only excludes the MLP layers that carry fine ornament — the bangles are named in 34 of 48 captions and still render grey. Not fixable by prompting.

  • Strict side profile and extreme expressions are unreliable. "Strict side profile view" and "furious snarling" both tend to come back as three-quarter view with a mild expression.

  • Not a 100% screen-accurate cel look — see "Read this first" above.

  • Do not prompt "photorealistic badger." Keep her cartoon, make the photo realistic.

Training

Base Krea 2 (inference on Turbo)

Network LoRA, rank 64 / alpha 64, attention layers only

Dataset 48 images, hand-captioned

Source Sonic Boom season 1, 1080p BluRay frames

Steps 1,440 (30 epochs), lr 1e-4, adamw8bit

Trainer musubi-tuner

Every frame was checked by hand for clipped framing, motion blur, wrong character and duplicates, and every caption names the outfit explicitly so the trigger carries identity rather than clothing.