dazzle-generator / README.md
huketo's picture
docs: name Bivri as the developer of Dazzle
6d36792 verified
|
Raw History Blame Contribute Delete
2.34 kB
metadata
license: apache-2.0
library_name: dazzle
tags:
  - image-protection
  - style-mimicry
  - protective-pattern

Dazzle generator v0.1.0

A 3,667,619-parameter fully convolutional U-Net that adds a bounded protective pattern to an illustration. The pattern is meant to reduce how much of the artist's style an SDXL LoRA learns when the image is used for unauthorized style-mimicry fine-tuning. One forward pass, any resolution whose sides are multiples of 8 (the library pads and crops other sizes).

Dazzle is developed by Bivri (비브리), an illustration community built on the AT Protocol where artists share artwork and exchange commissions.

  • Code: https://github.com/bivri/dazzle (tag v0.1.0, built from 0.1.0+73cb8a73537efaf4e9e1893967d9834713809501)
  • Output: delta = epsilon * tanh(G(x)), default bound 20.0 of 255 levels per channel

Use

from PIL import Image
from dazzle import DazzleGenerator, protect_image

gen = DazzleGenerator.from_pretrained("bivri/dazzle-generator", revision="v0.1.0")
protect_image(Image.open("art.png"), gen).save("art.protected.png")

Evaluation

Evaluation run generator-target-texture: for each of four synthetic illustration styles, one rank-16 SDXL LoRA (Illustrious-XL v2.0, 1,500 steps, kohya-ss/sd-scripts) was trained on the protected set, optionally after a transformation filter, and 40 renders were compared with an unprotected LoRA's renders in CSD style space. E/F is the fraction of the unprotected LoRA's style-specific gain that the protected set removed: 1 removes all of it, 0 none. LPIPS is measured at native resolution against the original images.

Training set given to the mimicry LoRA LPIPS E/F mean [95% CI]
protected (s1) 0.066 +0.110 [+0.039, +0.175]
protected (s4) 0.066 -0.214 [-0.316, -0.127]
protected+vae (s1) 0.081 +0.142 [+0.063, +0.210]
protected+vae (s4) 0.081 -0.268 [-0.371, -0.177]

Limitations

  • Results are from four synthetic styles, one LoRA recipe and one training seed per condition. CSD similarity is not a human judgment.
  • A transformation filter or a differently configured trainer can reduce the effect. The pattern does not make mimicry impossible.
  • The pattern is visible on close inspection of flat regions at the default bound.