<p>My attempt at a ControlNet for generating <strong>artistic yet scannable QR codes</strong> using SDXL.</p><p>Don't expect miracles, some images won't be scannable, or may require a specific angle / multiple attempts.</p><h2 id="how-it-works">How it works</h2><p>This model was <strong>not trained on real QR codes</strong>. Instead, it was trained on artistic images paired with binary grids created through adaptive thresholding (the same method QR scanners use to read codes).</p><ul><li><p><strong>Base model</strong>: SDXL 1.0</p></li><li><p><strong>Training data</strong>: ~1,700 SDXL-generated artistic images with dataset expansion tricks (via horizontal flips and rotations)</p></li><li><p><strong>Conditioning</strong>: 29x29 module grids</p></li><li><p><strong>Custom loss</strong>: Grid preservation loss on top of standard MSE to directly optimize for grid fidelity</p></li></ul><h2 id="how-to-use">How to use</h2><h3 id="1.-generate-your-qr-code">1. Generate your QR code</h3><p>I used <a target="_blank" rel="ugc" href="https://qrcode.antfu.me/">Anthony Fu's QR code generator</a>.</p><p>Recommended settings:</p><ul><li><p>Keep the size small. The model was trained on 29-module grids.</p></li><li><p>Keep the URL's short. Less data = better results.</p></li></ul><p>Use the generated QR code as ControlNet input.</p><h3 id="2.-controlnet-settings">2. ControlNet settings</h3><ul><li><p><strong>ControlNet strength</strong>: &gt; 2 (around 2.6 worked well for me)</p></li><li><p><strong>Prompts matter</strong>: Help the model create something with a lot of high contrast patterns.</p></li></ul><h3 id="3.-prompting-tips">3. Prompting tips</h3><p>This model responds best to high-contrast styles with sharp edges, like anime, geometric or pixel-like art patterns. Avoid prompts that fight the grid structure to much.</p><h2 id="support">Support</h2><p>Like this model? Consider <a target="_blank" rel="ugc" href="https://ko-fi.com/willempje2">buying me a coffee</a>. If I can break even on production costs, I'd like to train an improved version.</p>

Files with SHA256: f9bf277f5a7ab2414433f8cdd1d8a07793a1e14a4126a483fb5df4937c97f753