linoyts/Krea2-emoji-LoRA

Krea 2 Emoji Style LoRA — linoyts/emoji-krea2-r32-default

A modern 3D emoji style LoRA for Krea 2 — a fresher alternative to older SDXL-era emoji LoRAs. Trained on RAW, run on Turbo.

<Gallery />

Trigger phrase

End your prompt with `3d emoji`. Describe the subject in plain words and the LoRA renders it as a glossy, isolated-on-white 3D emoji / avatar.

Usage with 🧨 diffusers

Krea 2 support is on diffusers main, so install from source:

pip install -U git+https://github.com/huggingface/diffusers

import torch
from diffusers import Krea2Pipeline

# Load the LoRA onto Krea 2 Turbo (the 8-step distilled inference model).
# The LoRA was trained on Krea 2 RAW but expresses strongly on Turbo.
pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Turbo", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("linoyts/emoji-krea2-r32-default")

# Turbo recipe: 8 steps, no classifier-free guidance (guidance_scale=0.0).
# Resolution-aware timestep shifting is applied automatically by the scheduler — no `mu` needed.
image = pipe(
    "a wise wizard with a long white beard and a tall blue pointed hat, 3d emoji",
    num_inference_steps=8,
    guidance_scale=0.0,
    height=1024,
    width=1024,
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("emoji.png")

For weighting / merging / fusing LoRAs, see the diffusers LoRA docs.

Training details

• —Base: krea/Krea-2-Raw (validation on krea/Krea-2-Turbo), via the Krea 2 DreamBooth LoRA trainer.
• —Dataset: linoyts/3d-emoji-1024 — 306 emoji upscaled to 1024px and re-captioned (subject description + 3d emoji anchor), following the guide's style-LoRA captioning advice.
• —LoRA: rank 32 / alpha 32, full default layer set.
• —Optimization: lr 3e-4 constant, 8-bit AdamW, 1600 steps, 1024px, bf16, gradient checkpointing + cached latents.

This was the best of a 3-way sweep (rank 16 / 32 / 64) — rank 32 gave the most consistent, cleanest results.

License / provenance

The training data derives from Apple-style emoji artwork and this LoRA reproduces that aesthetic. Provided for research and experimentation — respect the original artwork's IP. Marked other.