loraspectra: Fast zero-dependency Rust CLI to audit, shrink (-60% to -75%), and orthogonally merge SD/SDXL/Flux LoRA .safetensors

[SalyyS1]

v0.6.0 Update - Try Live in Browser (Drop any .safetensors file locally, zero upload): https://salyys1.github.io/loraspectra/

• Inspect HuggingFace Hub directly (no manual download): loraspectra audit hf://org/repo
• Automated Merge Advisor: loraspectra interference a.safetensors b.safetensors recommends linear, ties, or ortho-svd + per-block --layer-weights.
• GitHub & Prebuilt Binaries: https://github.com/SalyyS1/loraspectra

Hi everyone,

We built loraspectra, a fast, zero-dependency Rust CLI tool for inspecting, auditing, resizing, extracting, and merging .safetensors LoRA and DoRA checkpoints on CPU in milliseconds.

What you can do with loraspectra

1. Audit any .safetensors LoRA in milliseconds (loraspectra audit):
   • Automatically identifies the base model family (SD 1.5, SDXL, Flux.1/Flux.2, etc.), rank distribution, dtype, and .alpha / DoRA magnitude tensors without needing a GPU or Python environment.
   • Computes exact singular values and flags DEAD_LAYER (untrained zero blocks) or over-parameterized blocks (OVER_PARAM).
2. Shrink oversized LoRAs by 60%-75% with zero visual degradation (loraspectra compress):
   • Uses exact Eckart-Young SVD (--energy 0.99) or Marchenko-Pastur noise removal (--policy mp) and casts to fp16 / bf16.
3. Merge multiple Character + Style LoRAs without interference (loraspectra merge --mode ortho-svd):
   • Projects secondary LoRAs onto the orthogonal complement of primary attention directions so character faces and style details do not wash each other out when combined.

Repository & prebuilt binary: https://github.com/SalyyS1/loraspectra

If you find this useful for managing your Stable Diffusion LoRA collection, a star on SalyyS1/loraspectra is much appreciated!