Z(oroj)L(indong)V(alentine)

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.

WHAT'S INSIDE

lindong3_IL01: 0.75 (Style)

Valentine_Eisenberg_spikes_IL: 0.3 (Style)

zoroj_ill11: 0.6 (Style)

NOOB vp1 Detailer: 1.0 (Detail enhancer)

All scored against the IL01 base checkpoint for compression.

VARIANTS

fro_ckpt thr=-2.5 46 MB -- Most retained, closest to unpruned

fro_ckpt thr=-2.2 17 MB -- Good balance, my go-to pick

fro_ckpt thr=-1.8 3.0 MB -- Very aggressive, still holds up

fro+spn thr=-2.0 31 MB -- 60% fro + 40% spn, middle ground

spn_ckpt thr=-1.5 40 MB -- Conservative, detailer-friendly

spn_ckpt thr=-1.3 21 MB -- Balanced, detailer-friendly

Naming: 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.

HOW IT WORKS

Instead 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.

COMPATIBILITY NOTE

The 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.

USAGE

Load 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.

CREDITS

@Queria https://civarchive.com/articles/5381/resizing-sdxl-loras-in-seconds-instead-of-minutes

QR+SVD LoRA compression algorithm (MIT-licensed open source)

sd-scripts LoRA training infrastructure

Original LoRA authors (see individual model pages)