Compress LoRAs by 65‑94%. Load 10‑20× more. Unlock exponential knowledge combinations.
🚀 Quick Start • 💼 Pro $49 • 🏢 Studio $149 • 📖 Full Docs
LoRA files are 144–487 MB each. You can load maybe 2–3 at a time before your VRAM is gone. That means your model is only accessing a fraction of the specialized knowledge available to it.
LoRA Lens fixes this. Compress your LoRAs by 65–94%, load 10–20 simultaneously on the same hardware, and unlock exponentially more knowledge combinations:
3 LoRAs loaded → 7 combinations
10 LoRAs loaded → 1,023 combinations
15 LoRAs loaded → 32,767 combinations
Same GPU. Same model. Thousands of times more capability.
Tested on real, publicly available LoRAs. Verify these yourself — download links below.
| Format | Original | After Rank Opt | After Quantization | Total Reduction |
|---|---|---|---|---|
| SD 1.5 | 144 MB | 18 MB | 9 MB | 93.8% |
| SDXL | 144 MB | 16 MB | 8 MB | 94.4% |
| FLUX | 487 MB | 175 MB | 87 MB | 82.1% |
| Metric | SD 1.5 | SDXL | FLUX |
|---|---|---|---|
| Variance Retained | 99.2% | 99.4% | 99.1% |
| SNR | 48.2 dB | 51.7 dB | 46.8 dB |
| MSE | 0.0012 | 0.0008 | 0.0015 |
| LoRA | Original | Compressed | Reduction |
|---|---|---|---|
| flux_koda_style | 342.0 MB | 1.4 MB | 99.6% |
| flux_anime_style | 42.8 MB | 1.9 MB | 95.6% |
| dmd2_sdxl_4step | 750.9 MB | 154.3 MB | 79.4% |
| hypersd_sdxl_2step | 750.9 MB | 186.0 MB | 75.2% |
| hypersd_sdxl_1step | 750.9 MB | 193.9 MB | 74.2% |
| hypersd_sdxl_8step | 750.9 MB | 239.6 MB | 68.1% |
| hypersd_sdxl_4step | 750.9 MB | 245.4 MB | 67.3% |
| lcm_lora_sd15 | 128.4 MB | 71.0 MB | 44.7% |
| hypersd_sd15_4step | 256.7 MB | 142.8 MB | 44.4% |
| hypersd_sd15_8step | 256.7 MB | 143.1 MB | 44.3% |
| TOTAL | 4,780.9 MB | 1,379.3 MB | 71.2% |
All 10 extract with max weight difference < 0.001. Full database as .loradb: 727.9 MB (47% additional reduction).
🔗 Download source LoRAs to verify
- Hyper-SD (ByteDance): https://huggingface.co/ByteDance/Hyper-SD
- DMD2 (tianweiy): https://huggingface.co/tianweiy/DMD2
- LCM-LoRA SD1.5: https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
- FLUX Style LoRAs: Available on CivitAI
Singular Value Decomposition identifies and removes unused dimensions in each layer's weight matrices. Most LoRAs are trained at higher ranks than they need — LoRA Lens finds the optimal rank automatically.
→ 30–90% reduction depending on format. Quality retention: 99%+.
Converts BFloat16/Float16 weights to Int8 with per-tensor symmetric scale factors. Designed specifically for visual model LoRAs, not adapted from LLM quantization.
→ Additional 50% reduction on top of Stage 1.
A new single-file format that stores collections of LoRAs using differential compression — only the weight deltas between LoRAs are stored. Related LoRAs (character variants, style series) compress dramatically:
Traditional: .loradb:
├─ character_1.safetensors 144 MB ├─ Base (compressed) 20 MB
├─ character_2.safetensors 144 MB ├─ Diff 1 (delta only) 2 MB
├─ character_3.safetensors 144 MB ├─ Diff 2 (delta only) 3 MB
└─ Total: 432 MB └─ Total: 25 MB (94% smaller)
| Collection | Individual | As .loradb | Reduction |
|---|---|---|---|
| 50 character LoRAs | 7.2 GB | 380 MB | 94.7% |
| 100 style LoRAs | 14.4 GB | 890 MB | 93.8% |
| 20 lighting LoRAs | 2.88 GB | 145 MB | 95.0% |
Reconstruct any individual LoRA on-demand in milliseconds. Compatible with ComfyUI, A1111, and all standard tools after extraction.
📐 .loradb format specification
.loradb File Structure:
├─ Header (magic bytes: 'LORA', version, count, metadata length)
├─ Metadata (JSON: collection info, LoRA manifest, offsets)
├─ Base LoRA (first LoRA, fully compressed)
├─ Differential LoRA #2 (weight deltas from base, sparse format)
├─ Differential LoRA #3 (weight deltas from base)
└─ ...
Reconstruction: base + diff_N = original LoRA_N
git clone https://github.com/intuitivation/LoRA-Lens.git
cd LoRA-Lens
pip install -r requirements.txt
python run_lens.pyOpens automatically at http://localhost:8501. Windows users: just run launch_lens.bat.
Requirements: Python 3.8+ · 8 GB RAM (16 GB recommended) · Works on CPU, faster with GPU
Try it now: The repo includes demo_collection.loradb — a mini database with 2 FLUX LoRAs (3.3 MB) you can extract and inspect immediately.
See QUICKSTART.md for a full walkthrough with screenshots.
| Tab | What It Does |
|---|---|
| Dashboard | Real-time LoRA analysis — health score, efficiency metrics, format auto-detection |
| Analytics | Layer-by-layer weight distributions, correlation heatmaps, sparsity visualization |
| 3D Topology | UMAP projection of weight patterns with interactive cluster exploration |
| Conflict Scanner | Test two LoRAs for layer conflicts before merging — get ratio recommendations |
| AI Consultant | Ask questions about your LoRA in plain English, get optimization advice |
| Optimize | One-click SVD rank optimization with batch processing support |
| Surgery (Pro) | 8-bit/4-bit quantization, Ultra Compress mode, real-time quality metrics |
| Export | Download optimized LoRAs, create and extract .loradb collections |
| 🆓 Free | 💼 Pro · $49 | 🏢 Studio · $149 | |
|---|---|---|---|
| SVD rank optimization | ✅ | ✅ | ✅ |
| All formats (SD 1.5 / SDXL / FLUX) | ✅ | ✅ | ✅ |
| Analysis, visualizations, AI consultant | ✅ | ✅ | ✅ |
| Conflict detection & batch processing | ✅ | ✅ | ✅ |
| 8-bit / 4-bit quantization | — | ✅ | ✅ |
| Ultra Compress (rank + quant combined) | — | ✅ | ✅ |
| Real-time quality metrics (SNR, MSE, MAE) | — | ✅ | ✅ |
| .loradb creation | Up to 5 | Up to 50 | Unlimited |
| .loradb extraction | ✅ Unlimited | ✅ Unlimited | ✅ Unlimited |
| Commercial use | — | ✅ | ✅ |
| Sell/distribute .loradb files | — | — | ✅ |
| Users per license | 1 | 1–10 | 1–25 |
| Priority email support | — | 48hr | 24hr |
One-time payment. Lifetime updates. No subscription.
| Version | What's Coming |
|---|---|
| v1.7 | Batch .loradb creation · marketplace integration · collection management UI |
| v2.0 | REST API + Python SDK · cloud processing · collaborative collections |
| Future | Automatic LoRA categorization · version control · direct tool integrations |
Have a feature idea? Open an issue.
Bugs & feature requests: GitHub Issues Pro/Studio priority support: jonwright.24@gmail.com
Contributions welcome — bug fixes, documentation, feature suggestions. Open an issue to discuss before submitting major changes.
MIT for personal and educational use. Commercial license required for business use. Studio license required to sell .loradb files. Full terms in LICENSE and COMMERCIAL_LICENSE.md.
LoRA Lens is built and maintained by Jon Wright. If this tool helps your workflow:
⭐ Star this repo — helps others find it ☕ Buy me a coffee — support development 💼 Buy a license — funds continued development 📣 Share your results — post compression wins on Reddit, CivitAI, X
LoRA Lens v1.6 · Made with ❤️ for the AI community, Zoey, and Robl ✨ ·
