krea2-style-2p5d-brushstroke

Krea 2 Style LoRA — 2.5D Solid Brush (Knife Strokes)

A style LoRA trained on Krea 2, capturing a flat-shaded 2.5D concept-art brushwork style (front/back character turnaround sheets, knife-stroke rendering) from a 25-image curated dataset.

Details

• Base model trained on: krea/Krea-2-Raw (LoRA training requires the undistilled Raw checkpoint)
• Inference model: apply this LoRA to krea/Krea-2-Turbo for fast 8-step generation
• Trigger token: kbrsh2d
• Rank: 32
• Steps: 1000 (checkpoints saved every 200 steps: 200/400/600/800/1000)
• Trainer: DiffSynth-Studio

Usage

import torch
from diffsynth.pipelines.krea2 import Krea2Pipeline, ModelConfig

pipe = Krea2Pipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="krea/Krea-2-Turbo", origin_file_pattern="turbo.safetensors"),
        ModelConfig(model_id="Qwen/Qwen3-VL-4B-Instruct", origin_file_pattern="*.safetensors"),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-VL-4B-Instruct", origin_file_pattern=""),
)
pipe.load_lora(pipe.dit, "checkpoints/step-1000.safetensors")

prompt = "kbrsh2d, a knight in ornate plate armor holding a sword and shield, standing on a plain background"
image = pipe(prompt, seed=42, height=1024, width=1024, num_inference_steps=8, cfg_scale=1, mu=1.15)
image.save("image.png")

Checkpoints

Dataset

Trained on 25 curated concept-art images (front/back character turnaround sheets in a flat 2.5D painterly brush style), captioned content-only (no style words) with the layout described explicitly to avoid it bleeding into the learned style.

Sample outputs

See samples/ for evaluation renders across checkpoints (Turbo + LoRA, fixed seed/prompt per checkpoint).

Evaluation Findings

Tested all 3 later checkpoints (600/800/1000) on novel subjects not in the training set (astronaut, knight, chef, wolf) using Krea-2-Turbo + LoRA, 8-step inference.

• Style transfer: The flat, painterly 2.5D brushstroke style transfers well and generalizes to subjects outside the training set at all three checkpoints.
• Overfitting signal: Starting at step-800, and more visibly at step-1000, the model reproduces a small red "K" logo watermark that was baked into a few training images (an artifact of the source dataset that wasn't cropped out). step-600 does not exhibit this artifact.
• Recommendation: Use checkpoints/step-600.safetensors as the primary checkpoint — strong style transfer without the watermark leakage. If you need a stronger stylization and can tolerate manually cropping out the watermark in post, step-800 is the next option.
• Next step for a cleaner LoRA: identify and crop/inpaint the watermarked source image(s) in the dataset, then retrain — this should let higher step counts (800-1000) be used safely without the artifact.

See samples/ for the full set of comparison renders (3 checkpoints × 4 prompts).

Additional samples (step-600, batch 2)

Generalization test with streetwear / skate-culture prompts, plus a "recall" test (prompting with training-set-style captions to check memorization):

• samples/swag_p0_hoodie_chain.png — oversized hoodie + chain, style transfers well to new subject matter
• samples/swag_p1_skater_kickflip.png — skater mid-kickflip
• samples/swag_p2_tracksuit_durag.png — tracksuit + durag
• samples/recall_p4_loot_hoodie.png — recall prompt closely reproducing a training subject's silhouette
• samples/recall_p5_jumpsuit.png — recall prompt closely reproducing a training subject's silhouette

Two additional outputs from this batch were withheld from this repo because they rendered visually close to existing copyrighted characters (an original "hero" prompt read as Spider-Man-adjacent, and a recall prompt reproduced a Batman-like design present in the training set) — flagging this as a known risk when prompting near comic/superhero archetypes with this LoRA.