Forge Neo Krea2 LoRA Tools + SVD Resizer
A standalone Forge Neo extension for inspecting, structurally stripping, and SVD-rank-resizing Krea2 LoRAs directly from the WebUI.
It adds a new Krea2 LoRA Tools tab to the WebUI, with two modes:
- SVD Resize — analyze a LoRA's per-layer SVD energy curve, get a recommended rank for a given quality target, and re-save the LoRA at a smaller rank.
- Structural Stripper — remove specific groups of Krea2 diffusion-block tensors (by profile) to shrink a LoRA, with a dry-run mode and a risk-threshold safety check.
Structural analysis reads only tensor shape/dtype metadata from the safetensors header wherever possible, while SVD analysis loads the A/B tensor data needed for the computation.
Features
SVD Resize
The SVD rank-resizing implementation in this extension is a separate approach
from the structural stripping method above. It performs per-layer SVD
compression of LoRA A/B pairs and does not remove diffusion blocks by profile.
- Detects lora_A / lora_B pairs and validates that the file uses a supported Krea2 key layout diffusion_model.blocks.*, diffusion_model.txtfusion.*, transformer.transformer_blocks.*, transformer.text_fusion.*).
- Analyze & recommend rank: computes the retained-energy curve for each A/B pair using a QR + small-SVD approach, then suggests a target rank for Aggressive (95%), Balanced (98%), and Conservative (99%) quality targets — without writing any file.
- SVD Resize LoRA: truncates every pair to the chosen target rank and saves a new ..._SVD_r<rank>.safetensors file.
- Configurable alpha handling and compute device auto / cpu / cuda).
- Dry-run mode shows the estimated size reduction before you commit to writing a file.
Structural Stripper
The original Krea2 LoRA stripping approach was developed by Winnougan
That project credits Puppet_Master on Civitai as the original source
of the technique/code for reducing Krea 2 LoRA file sizes by removing
diffusion/DIT weights while retaining the text-fusion layers.
Our Structural Stripper is an expanded Forge Neo implementation inspired by
that approach, with additional profiles, dry-run analysis, risk thresholds,
and Forge Neo UI integration.
Original source: Puppet_Master on Civitai Red
- Four built-in profiles:
- Max (txtfusion only) — strips all diffusion-block LoRA weights, keeps only text-fusion tensors.
- Balanced (keep 50% diffusion blocks)
- Compact (keep 25% diffusion blocks)
- Light (aux only) — strips only auxiliary tensors first, last_linear, tmlp_, tproj_1, etc.).
- Analyze LoRA: reports total size, per-group tensor/byte breakdown, and dtype counts.
- Strip LoRA: writes a new, smaller ..._<Profile>.safetensors file. Dry-run mode is available, and a configurable risk threshold flags profiles that would strip an unusually large share of the file.
General
- Live-streaming log output and a progress bar for long-running operations, based on parsed N/M progress information from the processing logs.
- Won't run processing operations while a generation job is active in Forge Neo, to avoid interfering with the model/LoRA state or contending for GPU/VRAM mid-generation.
- Scans all of Forge's known LoRA directories models/Lora, models/lora, models/LoRA, models/loras, plus any --lora-dir--lora-dirs set on the command line).
Requirements
safetensors
torch
gradio
These dependencies are normally already provided by a standard Forge Neo / Stable Diffusion WebUI Forge installation, so no extra installation is usually required.
Installation
1. Download or clone this repository into your Forge Neo extensions/ folder:
cd <your-forge-neo>/extensions git clone https://github.com/dbodik-git/forge-neo-krea2-lora-tools.gitReplace <your-forge-neo> with your Forge Neo installation directory.
(Or extract the release ZIP into extensions/ directly.)
2. Restart Forge Neo (or use Reload UI).
3. A new Krea2 LoRA Tools tab will appear in the WebUI.
Usage
1. Open the Krea2 LoRA Tools tab.
2. Pick a LoRA from the dropdown (click 🔄 Refresh if you just added a file).
3. Choose a mode:
- SVD Resize — click Analyze & recommend rank first to see the energy curve and a suggested rank, then set SVD target rank and click SVD Resize LoRA (or enable SVD dry run to preview without writing a file).
- Structural Stripper — pick a Profile, click Analyze LoRA to review what would be removed, then Strip LoRA to write the reduced file (or enable Dry run first).
4. Progress and results are streamed into the log box below the buttons; output files are written alongside the source LoRA with a suffix describing what was done (e.g. _SVD_r16, _Max, _Balanced).
Notes & Limitations
- Only Krea2-layout LoRAs are supported for the SVD path — a file must expose matching lora_Alora_B pairs under one of the recognized key prefixes, or analysis/resize will report that no compatible pairs were found.
- SVD Resize builds the output tensors in memory before saving, so peak RAM/VRAM usage scales with the size of the generated output (plus whatever Forge already has resident). On memory-constrained systems, very large source LoRAs combined with a large loaded checkpoint may still hit a memory ceiling.
- This extension only processes local LoRA files already present on disk; it does not download or fetch anything from the network.
License
License MIT