<h2 id="z(oroj)l(indong)v(alentine)">Z(oroj)L(indong)V(alentine)</h2>An experimental SVD-compressed SDXL LoRA merging 3 style LoRAs + 1 detail enhancer into a single file. Reduced from ~700MB raw to 3-46MB via context-aware rank pruning -- keeps only what the base model can't already do. Results look decent to my eye, but this is a personal experiment; YMMV. All example images are raw outputs -- no high-res fix, no face restoration, no post-processing.<hr />WHAT'S INSIDE<a target="_blank" rel="ugc" href="https://civarchive.com/models/1021146/lindong-style-ilxl">lindong3_IL01</a>: 0.75 (Style)<a target="_blank" rel="ugc" href="https://civarchive.com/models/1217337/valentine-eisenberg-style-illustrious?modelVersionId=1374999">Valentine_Eisenberg_spikes_IL:</a> 0.3 (Style)<a target="_blank" rel="ugc" href="https://civarchive.com/models/1015562/zoroj-style-ilxlnai-xl">zoroj_ill11</a>: 0.6 (Style)<a target="_blank" rel="ugc" href="https://civarchive.com/models/987736/noobai-xl-detailer">NOOB vp1 Detailer</a>: 1.0 (Detail enhancer)All scored against the IL01 base checkpoint for compression.<hr />VARIANTSfro_ckpt thr=-2.5 46 MB -- Most retained, closest to unprunedfro_ckpt thr=-2.2 17 MB -- Good balance, my go-to pickfro_ckpt thr=-1.8 3.0 MB -- Very aggressive, still holds upfro+spn thr=-2.0 31 MB -- 60% fro + 40% spn, middle groundspn_ckpt thr=-1.5 40 MB -- Conservative, detailer-friendlyspn_ckpt thr=-1.3 21 MB -- Balanced, detailer-friendlyNaming: fro_ckpt = Frobenius norm vs checkpoint (total layer energy), spn_ckpt = Spectral norm vs checkpoint (strongest direction), ckpt = checkpoint (the base model).Thresholds were hand-picked so visual results are close across methods. All variants are lossy; subtle differences exist, but squeezing ~700MB of raw LoRa data this far is already a win.<hr /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/be18c646-236a-46ba-943e-073eb0724d14/original=true/be18c646-236a-46ba-943e-073eb0724d14.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ec876599-e3b9-4900-8ef4-da0a4fa595bf/original=true/ec876599-e3b9-4900-8ef4-da0a4fa595bf.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6041aec6-a203-499a-a359-f5522a6a6acb/original=true/6041aec6-a203-499a-a359-f5522a6a6acb.jpeg" /><hr />HOW IT WORKSInstead of giving every layer the same rank, each LoRA direction is compared against the base model's own strength at that layer. A direction only survives if it meaningfully exceeds what the base model already provides. Layers where the base model is strong get pruned harder; weaker layers keep more.The SVD step uses a QR decomposition trick to avoid building the full merged weight matrix, making it about 30x faster than traditional full-SVD LoRA resizers.<hr />COMPATIBILITY NOTEThe NOOB detailer used a Diffusers-based trainer with different key naming. A key mapper was built to bridge both schemes so all layers could be evaluated. Most of the detailer's extra resnet/conv layers were pruned by SVD, but its attention-layer contribution still makes a subtle difference in fine detail -- noticeable in side-by-side comparison, though the overall look is dominated by the three style LoRAs.<hr />USAGELoad with IL01 or any Illustrious-family SDXL checkpoint. Works with standard LoRA nodes in ComfyUI, Forge, or A1111. Trigger words from the original style LoRAs are preserved. The detailer applies globally -- no trigger needed.<hr />CREDITS@Queria <a target="_blank" rel="ugc" href="https://civarchive.com/articles/5381/resizing-sdxl-loras-in-seconds-instead-of-minutes">https://civarchive.com/articles/5381/resizing-sdxl-loras-in-seconds-instead-of-minutes</a>QR+SVD LoRA compression algorithm (MIT-licensed open source)sd-scripts LoRA training infrastructureOriginal LoRA authors (see individual model pages)

Files with SHA256: 30d92ef614ba3bbaf9dc811c767a5e2c8626f2144dc402856df7f059b5cb08c8