Effect of Dazzle on Unauthorized Style-Mimicry LoRA Fine-Tuning.
- Row 1 (
Unprotected LoRA): Generations from an SDXL LoRA fine-tuned on unprotected original illustrations.- Row 2 (
Dazzle-Protected LoRA): Generations from a LoRA fine-tuned on Dazzle-protected illustrations (E/F = +0.281,PSNR = 31.9 dB,LPIPS = 0.027).- Row 3 (
Dazzle + VAE Filter): Generations from a LoRA fine-tuned after passing Dazzle-protected images through an SDXL VAE encode-decode transformation filter (E/F = +0.151).
Named after Dazzle Camouflage (the high-contrast geometric ship patterns designed to disrupt optical rangefinders without concealing the vessel itself), Dazzle applies a bounded, non-invasive protective pattern to illustrations so that unauthorized LoRA fine-tuning on latent diffusion models (SDXL family) learns less of the artist's style.
Unlike general-purpose noise that leaves visible grain on flat anime skies, watercolor paper, or character skin, Dazzle shapes its protective signal using multi-band frequency bounds and collinear RGB exposure gating:
- Subculture & Lineart Preservation: A two-sided valley ink detector and 4-directional ink enclosure isolate interior shading while keeping thin lineart, bare character skin, sunset sky gradients, and white watercolor paper spotless.
- Zero Hue Shift: Collinear RGB exposure projection and uniform radial chroma attenuation prevent color-cast blotches or rainbow speckle in dark shadows.
- Single-Pass Generator & ONNX Support: A lightweight fully convolutional U-Net (
DazzleGenerator) emits the bounded protective patterndelta = epsilon * tanh(G(x))at native image resolution on GPU or CPU. - Transformation-Filter Resilience: Trained with aspect-ratio bucket resizing, JPEG compression, WebP, and SDXL VAE round trips in the loop.
Non-Invasive Visual Fidelity Across Illustration Styles (
1. Originalvs2. Dazzle-Protectedand3x Magnified Crops). Dazzle maintainsPSNR 34–45 dBandLPIPS 0.006–0.062across Flat Cel Anime, Pastel Watercolor Anime, 1990s Retro Cel Anime, and Painterly Illustrations.
pip install "bivri-dazzle @ git+https://github.com/bivri/dazzle"Requires Python 3.14+ and PyTorch 2.11+. Optional extras:
train: offline generator training (pip install "bivri-dazzle[train] @ git+https://github.com/bivri/dazzle")eval: end-to-end LoRA & CSD evaluation harness (pip install "bivri-dazzle[eval] @ git+https://github.com/bivri/dazzle")onnx: ONNX export (pip install "bivri-dazzle[onnx] @ git+https://github.com/bivri/dazzle")
# Protect one or more illustrations at native resolution (outputs PNG)
dazzle protect art.png more-art.jpg --out protected/
# Customize the per-channel intensity bound (in 8-bit levels, default: 20)
dazzle protect art.png --out protected/ --epsilon-levels 12
# Export the generator to ONNX with dynamic height and width axes
dazzle export-onnx --out dazzle.onnxfrom PIL import Image
from dazzle import DazzleGenerator, protect_image
gen = DazzleGenerator.from_pretrained("bivri/dazzle-generator")
protected = protect_image(Image.open("art.png"), gen)
protected.save("art.protected.png")Output is saved as lossless PNG. Re-saving as low-quality JPEG or heavy downscaling weakens the protective pattern.
- Single-Head Bounded Generator:
DazzleGeneratoris a fully convolutional U-Net that outputs a single tensordelta = epsilon * tanh(G(x))at the input illustration's native resolution (reflect-padded to a multiple of 8 and cropped back). - Composite Target-Texture Objective:
During training (
dazzle.training), the protected image is passed through differentiable transformation filters and encoded by the SDXL VAE (madebyollin/sdxl-vae-fp16-fix). The objective pulls the encoded latent and multi-scale feature statistics toward a composite target combining:- An Oklab chiaroscuro counter-bias image (flattening low-frequency contrast, inverting mid/high luminance bands, and radially attenuating chroma at
0°hue shift), and - A deterministic procedural CC0 anisotropic structural relief texture.
- An Oklab chiaroscuro counter-bias image (flattening low-frequency contrast, inverting mid/high luminance bands, and radially attenuating chroma at
- Multi-Band Non-Invasive Scalar Hinge Loss:
To keep flat backgrounds and delicate lineart untouched without adding auxiliary mask heads to the network, training regularizes
deltaacross low-, mid-, and high-frequency bands alongside LPIPS and total variation scalar losses.
Protection effectiveness is measured by E/F (effect_ratio in CSD style space): the fraction of an unprotected LoRA's style-specific gain removed by protection (1.0 = all style gain removed, 0.0 = no effect), evaluated across 20 held-out prompts against held-out style references (docs/evaluation.md).
| Evaluation Condition | Style | PSNR (dB) | LPIPS | Style Gain Removed (E/F [95% CI]) |
|---|---|---|---|---|
| Dazzle Multi-Band Target-Texture (48-Step) | Flat Cel Anime (s1) |
31.90 | 0.027 | +0.281 [+0.197, +0.355] |
| Dazzle + SDXL VAE Round-Trip Filter (48-Step) | Flat Cel Anime (s1) |
28.36 | 0.051 | +0.151 [+0.077, +0.224] |
Single-Pass DazzleGenerator + SDXL VAE Filter |
Flat Cel Anime (s1) |
28.77 | 0.081 | +0.142 [+0.063, +0.210] |
Single-Pass DazzleGenerator (Unfiltered) |
Flat Cel Anime (s1) |
30.66 | 0.066 | +0.110 [+0.039, +0.175] |
| Decoy-Latent Baseline | Flat Cel Anime (s1) |
31.85 | 0.108 | +0.138 [+0.084, +0.192] |
Visual Fidelity Probe Across Styles (00.png / 03.png) |
Max Bound ( |
PSNR (dB) | LPIPS |
|---|---|---|---|
Flat Cel Anime (s1) |
20–36 levels | 41.82 – 45.30 dB | 0.0060 – 0.0187 |
Pastel Watercolor Anime (s2) |
24–36 levels | 39.45 – 40.92 dB | 0.0217 – 0.0357 |
1990s Retro Cel Anime (s3) |
28–36 levels | 40.86 – 41.03 dB | 0.0148 – 0.0177 |
Painterly Illustration (s4) |
36 levels | 34.17 – 39.45 dB | 0.0266 – 0.0617 |
Scope & Limitations: Dazzle is an active research project targeting unauthorized style mimicry via LoRA fine-tuning on latent diffusion models. While flat cel and subculture lineart styles show substantial suppression at imperceptible distortion levels, richly textured painterly styles under strict visibility bounds and reference-based image editing remain open research areas.
| Path | Description |
|---|---|
src/dazzle/ |
Core library: DazzleGenerator, protect_image, training objective, CLI, and evaluation harness |
assets/ |
Vector SVG logos (dazzle-logo.svg, dazzle-mark.svg) and visual comparison showcases |
docs/ |
Evaluation protocol (docs/evaluation.md), design records, and contributor guides |
experiments/ |
Reproducible experiment records and JSON evaluation reports |
scripts/ |
Repository policy checks and Hugging Face Hub model publication script |
tests/ |
Unit and policy test suite run in CI |
See CONTRIBUTING.md and CONTEXT.md.
Copyright 2026 © LSware. Licensed under the Apache License, Version 2.0 — see LICENSE and NOTICE.