Barath/flux2-klein-yarn-art-lora

FLUX.2 [klein] 4B — Yarn Art style LoRA

A style LoRA for FLUX.2 klein base-4B that renders subjects as yarn / fiber art with dense stitch and thread texture.

Trigger word: Y4RN_ART — the style only activates when the trigger is in the prompt; without it, output stays photorealistic (no style bleed).

With trigger (fox)	With trigger (astronaut)	On distilled, 4 steps	No trigger

Usage

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-4B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("Barath/flux2-klein-yarn-art-lora")

img = pipe(
    "Y4RN_ART, a fox in a forest",
    num_inference_steps=28, guidance_scale=4.0,
).images[0]

The LoRA was trained on the undistilled base model but also transfers to the 4-step distilled FLUX.2-klein-4B:

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("Barath/flux2-klein-yarn-art-lora")
img = pipe("Y4RN_ART, a fox in a forest", num_inference_steps=4, guidance_scale=1.0).images[0]

Files

• klein4b_yarn_art_lora_v1.safetensors — final checkpoint (step 1800, recommended)
• klein4b_yarn_art_lora_v1_000001200.safetensors — step-1200 checkpoint (slightly lighter style)

Training

Trained with ostris/ai-toolkit on a single RTX 4090 (~31 min).

• Dataset: Norod78/Yarn-art-style (18 images, 1024px), captions rewritten to content-only with the Y4RN_ART trigger prefix
• LoRA rank 32 / alpha 32, lr 1e-4, AdamW 8-bit, 1800 steps, batch 1, bf16 with 8-bit quantization, flowmatch, EMA 0.99
• Multi-resolution buckets: 512 / 768 / 1024

Known limitation: still-life / tabletop scenes (e.g. "a cup of coffee on a wooden table") stay photorealistic — the small dataset contains creatures, figures and single objects, so coverage of such scenes is weak.