badaramoni/Krea-2-Realism-Fusion

Badaramoni Realism Fusion — Krea 2

A photorealism LoRA for Krea 2 Turbo, produced with Badaramoni Fusion: a format-agnostic, delta-space merge that combines multiple realism adapters — including ones stored in different low-rank formats — into a single portable LoRA.

No trigger word. Write a natural, descriptive sentence.

Why this isn't a normal merge

Off-the-shelf mergers (kohya, mergekit) only combine adapters that share the same low-rank format. Real-world realism adapters are a mix of classic LoRA and LoKR, which those tools cannot merge together. Badaramoni Fusion solves this by working in weight-delta space:

1. Reconstruct each adapter's true weight delta per module (ΔW = B·A for LoRA, ΔW = kron(w1, w2) for LoKR).
2. Normalize across differing key conventions so every module lines up in one canonical space.
3. Blend the deltas as a weighted sum (ΔW = Σ wᵢ·ΔWᵢ).
4. Re-factor the combined delta back into one low-rank LoRA via truncated SVD (rank 128), with a measured reconstruction-error budget.

The result is a single .safetensors that carries the combined realism direction of several source models, not a naive alpha-average. This is a many-to-one, lossy-by-design compression that keeps the dominant realism singular directions in a compact, portable adapter.

The merge method lives in universal_merge.py.

Samples

All images below were generated on Krea 2 Turbo with this LoRA at scale 1.3, 8 steps, guidance 1, μ 1.15. Prompts are plain descriptive sentences — no style keywords.

Sunflower field, golden hour	Kitchen, soft window light
Surfer at sunrise	Cafe counter, late light
Street musician at night	Rain, taxi, wet asphalt
Autumn park bench	Subway portrait

Recommended settings

• Base: krea/Krea-2-Turbo
• Steps: 8, guidance: 0–1, μ: 1.15
• LoRA scale: 1.0–1.3 (start at 1.3 for a stronger realism push)
• Resolution: 1024×1280 works well; short side 1024–1536
• No trigger word

Usage with 🤗 diffusers

Install diffusers from source (for Krea2Pipeline):

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

import torch
from diffusers import Krea2Pipeline

pipe = Krea2Pipeline.from_pretrained(
    "krea/Krea-2-Turbo", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("badaramoni/Krea-2-Realism-Fusion")

prompt = (
    "A grandmother rolling dough in a flour-dusted kitchen, "
    "soft window light, wrinkled hands and warm smile"
)
image = pipe(prompt, num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("fusion.png")

Usage with musubi-tuner

python src/musubi_tuner/krea2_generate_image.py \
  --dit turbo.safetensors \
  --vae qwen_image_vae.safetensors \
  --text_encoder qwen3vl_4b_bf16.safetensors \
  --lora_weight badaramoni_krea2_realism_fusion_r128_fp16.safetensors \
  --lora_multiplier 1.3 \
  --steps 8 --guidance_scale 1 --mu 1.15 \
  --prompt "A surfer walking out of the ocean at sunrise carrying a board"

Download

Weights: badaramoni_krea2_realism_fusion_r128_fp16.safetensors (rank 128, fp16, musubi lora_unet_* keys).

Method

Built with a custom format-agnostic merger:

• Mixed LoRA + LoKR adapters reconstructed into a shared ΔW space
• Equal-weight blend, then truncated SVD back to rank 128
• Per-module processing so large Kronecker deltas stay in VRAM budget
• Output is one portable LoRA — no multi-adapter stack required at inference

This release is a merge, not a from-scratch training run. The contribution is the fusion algorithm that makes mixed-format realism adapters usable as a single file.

License

Downstream use follows the Krea 2 Community License. You must have access to krea/Krea-2-Turbo to run this adapter.