Model Utility Toolkit

Fix a diffusers-format Anima LoRA for native ComfyUI loaders

Pull the preview image and embedded workflow out of a checkpoint

Checkpoint Downloader — bulk-fetch metadata for every checkpoint you own

What's actually inside a safetensors checkpoint

Compare two checkpoints before you merge them — and find out if merging is even worth it

Strip unwanted tensors out of a checkpoint file

Rename Checkpoint Keys — fix a checkpoint whose layer names don't match

A third input for when two-way blending isn't enough

Blend two full checkpoints with actual controls

ControlNet Info Loader — untangle a folder of identically-named ControlNets

ControlNet Downloader — build a local index of what each ControlNet actually does

Three checkpoints into one, where the blend adapts per layer

CWB's adaptive blend

Take the wheel from CWB presets — the config node behind every CWB merge

Merge three embeddings into one vector file

Merge two embeddings into one vector file

Merge 2 to 8 LoRAs in one go — the CWB LoRA blender that scales

Three LoRAs into one file, blended by consensus per layer

Merge two LoRAs into one file with CWB, instead of stacking them in a workflow

Three UNets/DiTs, one consensus-merged denoiser

Merge two diffusion models (UNet/DiT) with a per-layer consensus blend

Three text encoders, one consensus-merged CLIP

Merge two CLIP text encoders with a consensus blend

Two diffusion models, zero merges — see how similar they really are first

Resize a diffusion model's precision without re-downloading it

Diffusion Model Info Loader — see what a .safetensors file actually is

Diffusion Model Downloader — make sense of a folder full of quant files

Pull an adapter out of a base/finetune pair at a rank you choose

Adaptive-rank extraction that targets quality, not a fixed size

DoRA Extract (Knee Detection) — pull a DoRA out of two checkpoints, automatically

DoRA Extract (Quantile) — extract a DoRA sized by how much energy you want to keep

DoRA Extract (Ratio) — extract a DoRA by relative singular-value magnitude

DoRA Learned Extract (Fixed Rank) — SVD extraction refined by gradient descent

DoRA Learned Extract (Frobenius) — target energy retention, then refine it

DoRA Learned Extract (Knee Detection) — automatic rank, then gradient-refined

DoRA Learned Extract (Quantile) — percentage-targeted extraction, gradient-refined

DoRA Learned Extract (Ratio) — magnitude-threshold extraction, gradient-refined

Embedding Info Loader — read the label on a 30KB mystery file

Bulk-fetch metadata and previews for your textual inversions

Get Embedding Metadata & Keys — peek inside a .safetensors embedding

Compare two embeddings without guessing which tokens match

Prune Embedding Keys — strip tensors out of a .safetensors embedding

Rename Embedding Keys — fix tensor naming inside a .safetensors embedding

Merge Embeddings (3 Models) — blend three textual inversions into one

Merge Embeddings (2 Models) — blend two .safetensors embeddings into one

LoRA Extract (Fixed Rank) — pull a LoRA out of two checkpoints

LoRA Extract (Frobenius) — extract a LoRA sized by norm retention, not rank

LoRA Extract (Knee Detection) — auto-rank LoRA extraction from two checkpoints

LoRA Extract (Quantile) — extract a LoRA by cumulative singular-value mass

LoRA Extract (Ratio) — extract a LoRA by singular-value ratio threshold

Pull the preview image and embedded workflow out of a LoRA file

LoRA Downloader — bulk-fetch metadata and previews for your LoRA folder

Bake up to 8 LoRAs permanently into a checkpoint

See what a LoRA actually trained (and every tensor it has)

Two LoRAs, same concept? Check before you merge them

Combine up to 8 LoRAs into one file, ranks and naming resolved

LoRA Multi-Merge (DARE-Ties) — combine up to 8 LoRAs without them fighting

LoRA Multi-Merge (Enhanced DARE-Ties) — magnitude-aware LoRA merging for up to 8 LoRAs

Bake a LoRA's Alpha Into Its Weights (and Drop the Alpha Tensors)

Prune LoRA Keys — strip tensors out of a LoRA .safetensors file

Fix mismatched tensor names without retraining anything

LoRA Resize (Cumulative) — shrink a LoRA without hand-picking a rank

LoRA Resize (Fixed Rank) — shrink or resize an existing LoRA

LoRA Resize (Frobenius) — shrink a LoRA to a target fidelity, not a fixed rank

LoRA Resize (SV Ratio) — resize a LoRA by singular-value threshold

Three LoRA files through the generic weighted-merge engine

Blend two LoRA files with the same engine as the checkpoint mergers

Manual Path Downloader — fetch model metadata from any folder, not just loras/

Get Diffusion Model Metadata & Keys — open the hood on a .safetensors file

Prune Diffusion Model Keys — strip tensors out of a checkpoint file

Rename Diffusion Model Keys — fix tensor naming across checkpoint formats

Three-way blending for UNet-only diffusion models

Blend two UNet-only diffusion models

Extract a text-encoder DoRA at a fixed rank, no training required

Extract a text-encoder DoRA that keeps a target share of the energy

Let the singular values pick the rank

DoRA for text encoders, rank by quantile

DoRA for text encoders, rank by ratio

Turn a fine-tuned text encoder into a LoRA — no training run required

Keep X% of the Frobenius norm

The LoRA node that picks its own rank

Keep enough singular values to reach 90%

Keep the singular values that actually matter

Inspect a CLIP/T5 file before you merge or debug it

How different are they, really?

Prune Text Encoder Keys — strip tensors out of a standalone CLIP/T5/Qwen encoder file

When your text encoder ships with the wrong labels

Merging three text encoders you almost certainly don't have to

Blend two CLIP/T5 files at the weight level

VAE Info Loader — figure out which VAE file is which

VAE Downloader — fetch metadata for the VAE files you actually keep around