numzoo-flux2-klein-lora

NumZoo — FLUX.2-klein-4B Style LoRA 🐾

A cozy kawaii children's-book style LoRA for black-forest-labs/FLUX.2-klein-4B. It turns the base model's varied (sometimes photoreal) output into a consistent chibi aesthetic: big sparkling eyes, soft pastels, rich storybook backgrounds with warm fairy lights.

Built for NumZoo — a kids' mental-math app where correct answers earn a freshly generated cute-animal reward image.

Trigger word

Use NUMZOO at the start of the prompt. Append no text to stop the model rendering the trigger as a literal sign.

NUMZOO. A cute bunny in an enchanted mushroom forest, no text

Why it exists — before / after

Raw FLUX.2-klein-4B renders the same prompt in inconsistent styles (here, a photoreal panda). The LoRA pins it to the cozy NumZoo look.

Prompt: "a cute baby panda on a snowy mountain top"

Examples

Prompts (all prefixed with NUMZOO. and suffixed with , no text):

• A cute bunny in an enchanted mushroom forest
• A cute puppy and kitten in a cosy cottage garden
• A cute unicorn under a rainbow and on a glowing crescent moon
• A cute baby fox on a sunny beach with ocean waves
• A cute baby lion, baby tiger and bunny surrounded by sparkling stars

Usage

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("goumsss/numzoo-flux2-klein-lora", weight_name="numzoo_klein.safetensors")

image = pipe(
    prompt="NUMZOO. A cute bunny in an enchanted mushroom forest, no text",
    num_inference_steps=4,      # distilled klein: 4 steps
    guidance_scale=1.0,         # distilled klein: always 1.0
    height=512, width=512,
).images[0]
image.save("numzoo.png")

Trained on the base model (FLUX.2-klein-base-4B) and loaded on the distilled FLUX.2-klein-4B for fast 4-step inference — the BFL-recommended workflow.

Training

• Base: black-forest-labs/FLUX.2-klein-base-4B
• Data: 54 cozy scenes generated with Qwen-Image, captioned NUMZOO. <content> (content only — the style is learned from the trigger, not described in captions)
• Trainer: ostris/ai-toolkit on a Modal A100
• Recipe: LoRA rank 32, lr 1e-4, 1500 steps (checkpoint 1250 selected), 512–1024px buckets
• Coverage: 12 animals × 10 places, incl. 1–3 animal and 1–3 place combinations

License

Apache-2.0, matching the FLUX.2-klein-4B base model.