<h3 id="&quot;remove-the-semi-transparent-mosaic-overlay-and-faithfully-restore-the-underlying-details.&quot;">"Remove the semi-transparent mosaic overlay and faithfully restore the underlying details."</h3>This lightweight LoRA adds inverse alpha-blend restoration capability to Qwen-Image 2.1, enabling it to remove light to moderately heavy semi-transparent (opacity up to 0.9) mosaic overlays from images.We did not train it on specific body parts, viewing angles, or texture details, and it does not explicitly inject a new style or introduce new concepts. It is intended to handle semi-transparent mosaic overlays across a broad range of body parts/regions, texture details, cell sizes, and areas, although performance may vary depending on the input.It should be emphasized that Qwen-Image 2.1 already has some capability to remove mosaic occlusion. This adapter is intended to improve the precision of semi-transparent mosaic removal by encouraging the model to reconstruct occluded details primarily from texture and structure cues preserved under the semi-transparent overlay, rather than relying mainly on prior-driven guessing.<h2 id="usage">Usage</h2>To save processing time, keep the model focused, and avoid degrading quality in other regions, we recommend cropping the image to the approximate region of mosaic before processing. Once you have loaded the LoRA and configured your workflow as usual, spell:<pre><code>Remove the semi-transparent mosaic</code></pre>Recommend settings:<ul><li>Strength: 1.0</li><li>CFG: 1.0</li><li>Sampler: euler / simple</li><li>Steps: 10-20 (Base) | 4-8 (Turbo)</li><li>No trigger word</li></ul><h2 id="results-and-comparison">Results &amp; Comparison</h2>Visit the <a rel="ugc" href="https://huggingface.co/Schwi-AI/Qwen-Image-2.1-MosaicMate-LoRA#results">model page</a>.<h2 id="limits">Limits</h2>Since semi-transparent overlays still cause information loss, the adapter does not guarantee pixel-exact recovery of the original content. Results depend on the base model, overlay opacity, and input image; heavily degraded regions may not be faithfully recovered.This is our first LoRA with real practical value. It was trained using only Fizgig’s recommended preset settings and already produces satisfactory results. Any comments or suggestions are welcome.If you are aiming for pixel-perfect restoration and don’t mind performing some manual touch-ups, you might want to try our <a target="_blank" rel="ugc" href="https://github.com/Schwi-AI/MosaicMate">dedicated tool</a>.

Files with SHA256: 3eaf24cba51f70e5e19b5a42b6479e85d679ceb3f8fe578cb89023792d911897