import torch
from diffusers import Flux2KleinPipeline

# Load base model + LoRA
pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe.load_lora_weights("Limbicnation/pixel-art-lora")
pipe.to("cuda")

image = pipe(
    "pixel art sprite, a brave knight in shining armor, game asset, transparent background",
    num_inference_steps=4,
    guidance_scale=1.0,
    height=512,
    width=512,
).images[0]

image.save("knight.png")

ComfyUI

Download the LoRA weights:
bash
1huggingface-cli download Limbicnation/pixel-art-lora \ 2 pytorch_lora_weights.safetensors \ 3 --local-dir ./models/loras/
Load FLUX.2-klein-4B as the base checkpoint
Add a Load LoRA node, point to pytorch_lora_weights.safetensors
Set LoRA strength: 0.85 - 1.4 (tested range)
Use trigger words in your positive prompt

A sample workflow is included: comfyui_workflow.json

Sample Gallery

Preview	Prompt
Knight	pixel art sprite, a brave knight in shining armor holding a sword
Archer	pixel art sprite, an elven archer with bow and quiver
Orc	pixel art sprite, a fierce orc warrior with battle axe
Dragon	pixel art sprite, a fire-breathing dragon with red scales
Slime	pixel art sprite, a cute slime monster, blue and bouncy
Skeleton	pixel art sprite, a skeletal warrior with rusty sword
Unicorn	pixel art sprite, a majestic unicorn with rainbow mane
Vampire	pixel art sprite, a dark vampire lord with cape and fangs
Marine	pixel art sprite, a space marine in power armor
Cyberpunk	pixel art sprite, a cyberpunk hacker with neon implants
Robot	pixel art sprite, a cute robot companion with glowing eyes
Ninja	pixel art sprite, a ninja assassin in black outfit with katana

14/15 samples generated successfully (1 failed due to content filter)

Training Details

Parameter	Value
Base model	`black-forest-labs/FLUX.2-klein-4B`
LoRA rank	64
LoRA alpha	128
rsLoRA	Yes
Dtype	bfloat16
Steps	1000
Batch size	1
Gradient accumulation	4
Learning rate	1e-4
LR scheduler	Cosine with restarts
Optimizer	AdamW 8-bit
Resolution	512x512
Dataset	500 images (CC0 curated + synthetic)

Final loss: ~1.8 - 2.0 (stable convergence)

Full training config: config.yaml

Dataset & Provenance

Training dataset: Limbicnation/pixel-art-character (1,010 images)
License: All training images are CC0 (public domain) — no copyright restrictions
Provenance: Full manifest with source attribution available in the dataset repo
Legal analysis: Synthetic training data compliant for commercial use

This model was trained with full data provenance tracking. All source images are CC0 public domain, making the outputs safe for commercial game development, Steam releases, and B2B studio use.

Architecture Notes

FLUX.2-klein-4B uses a different architecture from FLUX.1:

Text encoder: Qwen3 (not CLIP+T5)
Pipeline class: Flux2KleinPipeline (not FluxPipeline)
VAE: AutoencoderKLFlux2
Distilled: 4-step inference with guidance scale 1.0

You must use diffusers >= 0.37.0.dev0 (install from git main branch). The released 0.36.x does not include Flux2KleinPipeline.

License

This LoRA adapter is released under Apache 2.0.

The base model (FLUX.2-klein-4B) is also Apache 2.0 licensed.

Citation

bibtex

@misc{pixel-art-lora-2026,
  title={Pixel Art Sprite LoRA for FLUX.2-klein-4B},
  author={Limbicnation},
  year={2026},
  url={https://huggingface.co/Limbicnation/pixel-art-lora}
}

Limbicnation/pixel-art-lora

Pixel Art Sprite LoRA for FLUX.2-klein-4B

Highlights

Quick Start

Trigger Words

Inference Parameters

Python (Diffusers)