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

A LoRA adapter trained on FLUX.2-klein-4B for generating pixel art character sprites. Optimized for game-ready assets with transparent backgrounds.

Highlights

• 4-step inference — FLUX.2-klein is distilled, so generation is fast
• 512x512 RGBA output with transparent backgrounds
• CC0 training data — 100% public domain, no copyright concerns
• Game-ready — designed for Godot, Unity, and other engines

Quick Start

Trigger Words

Always include in your prompt:

pixel art sprite, [your character description], game asset, transparent background

Style modifiers: 16-bit pixel art, 32-bit pixel art, chibi

Inference Parameters

Python (Diffusers)

Requires diffusers >= 0.37.0.dev0 (install from git main)

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

1. Download the LoRA weights:

   huggingface-cli download Limbicnation/pixel-art-lora \
       pytorch_lora_weights.safetensors \
       --local-dir ./models/loras/
   

2. Load FLUX.2-klein-4B as the base checkpoint
3. Add a Load LoRA node, point to pytorch_lora_weights.safetensors
4. Set LoRA strength: 0.85 - 1.4 (tested range)
5. Use trigger words in your positive prompt

A sample workflow is included: comfyui_workflow.json

Sample Gallery

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

Training Details

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

@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}
}

Links

• Model: Limbicnation/pixel-art-lora
• Dataset: Limbicnation/pixel-art-character
• Base model: black-forest-labs/FLUX.2-klein-4B
• Training code: Limbicnation/SpriteForge