LoRA

Nodes for loading LoRAs — by index, filtered, or as a blended pair — and for working with LoRA files themselves: rescaling, shrinking, and inspecting them.

At a glance

Node	Summary
LoRA Loader By Index	Loads a LoRA by its numeric position in a sorted directory listing — built for sweeping a collection across runs.
LoRA Loader Filtered	The built-in LoRA loader with a dropdown sortable by date, name, or size (newest-first by default).
LoRA Split Loader	Blends two LoRAs with a single balance slider — 0 is all A, 1 is all B, 0.5 is both at half strength.
LoRA List Directory	Lists the LoRA files in a folder plus a count — a companion to index-based loading.
LoRA Scale & Save	Bakes a strength change into a LoRA (via alpha or weights) and saves it as a new file.
LoRA Truncate Rank	Shrinks an SVD-extracted LoRA by slicing it to a lower rank — fast, no re-decomposition.
LoRA Metadata	Reads a LoRA's embedded training metadata: trigger words, base model, rank, tags, and more.
Safetensors Metadata	Generic safetensors inspector for any model type — metadata plus tensor structure.

---

Loading

LoRA Loader By Index

Loads a LoRA chosen by its numeric position in an alphabetically sorted directory listing, wrapping past the end — built for sweeping a whole collection across queued runs.

Point it at the loras folder (or a subfolder) and increment index each run to iterate every LoRA in turn. It searches every configured LoRA path, including extra_model_paths.yaml entries, and wraps the index modulo the file count instead of erroring — with a console banner when a full pass completes.

Parameter	Type	Default	Description
model	MODEL	—	Base model to apply the LoRA to.
subdirectory	STRING	""	Subfolder relative to a LoRA base path (can be nested, e.g. wan/characters). Blank = the base folder itself.
recursive	BOOLEAN	false	Include nested subfolders (files listed by relative path).
index	INT	0	Zero-based position in the sorted list. Wraps modulo the file count.
strength_model	FLOAT	1.0	LoRA strength on the model weights (-20–20).
strength_clip	FLOAT	1.0	LoRA strength on CLIP; only used when clip is connected.
clip	CLIP	(optional)	Optional CLIP to also patch.

Outputs: model, clip (or None if not connected), lora_name, total_loras, actual_index (post-wrap), wrapped (true if the index wrapped).

File types: .safetensors, .pt, .bin, .ckpt. Raises an error if the directory can't be found or contains no LoRA files.

LoRA Loader Filtered

A faithful clone of the built-in Load LoRA node whose dropdown can be sorted by modification date, name, or size — newest-first by default.

The sort is applied client-side in the picker, so grabbing the LoRA you just downloaded or trained is one click. Text search is the combo widget's native type-to-filter. Loading behavior matches the built-in node exactly, including the tensor cache and the zero-strength short-circuit.

Parameter	Type	Default	Description
model	MODEL	—	Model to patch.
clip	CLIP	—	CLIP to patch.
lora_name	choice	—	The LoRA file; the dropdown is reordered per the sort controls.
strength_model	FLOAT	1.0	Model strength.
strength_clip	FLOAT	1.0	CLIP strength.
sort_by	choice	date_modified	date_modified, name, or size — display order only.
sort_order	choice	descending	Sort direction — display order only.

Outputs: model, clip.

If both strengths are 0.0 the inputs pass through untouched with no file load. File dates/sizes come from a small server endpoint and are cached in the browser, refreshed only when new files appear.

LoRA Split Loader

Loads two LoRAs — A and B — and blends between them with a single balance slider: 0.0 applies A at full strength, 1.0 applies B, 0.5 both at half.

The strengths are A = (1 − balance) × strength_multiplier and B = balance × strength_multiplier, applied to the model (and CLIP when connected). Each row has its own enable dot, and clicking a row's name opens a searchable picker — multi-term substring filtering plus A-Z / Newest sorting. Picked LoRAs show their top trigger words as chips under the row (click a chip to copy it), and each row displays a live read-only preview of the strength the slider currently gives it.

balance is a plain FLOAT widget, so it can be converted to an input and driven by another node — for example Float Lerp to animate a morph from one LoRA to the other across a batch.

Parameter	Type	Default	Description
model	MODEL	—	Model to patch.
clip	CLIP	(optional)	Also patched at the same strengths when connected.
enabled	BOOLEAN	true	Off: model and CLIP pass through unchanged and no triggers are emitted.
balance	FLOAT	0.5	The A↔B slider: 0.0 = A only, 1.0 = B only.
strength_multiplier	FLOAT	1.0	Scales both final strengths (-10–10); 0 disables both.
lora_a, lora_b	row	—	The two rows: enable dot, LoRA picker, trigger chips, effective-strength readout.

Outputs: model, clip, triggers (the top trigger words of every applied LoRA, comma-joined and deduped — ranked from embedded training metadata, the same extraction as LoRA Metadata).

A side whose resolved strength is 0 — the slider at an extreme, or its row toggled off — is skipped entirely, including its trigger words; the other side keeps its slider-derived strength rather than renormalizing to full. Selecting the same LoRA in both slots loads the file once and applies it twice (the strengths sum). A row that references a missing file raises an error instead of silently skipping it.

LoRA List Directory

Lists the LoRA files in a directory as a newline-separated string, plus a count.

A companion to LoRA Loader By Index: check what's in a folder and how many items there are before setting up index-based iteration. Only the first configured LoRA path is searched.

Parameter	Type	Default	Description
subdirectory	STRING	""	Subfolder relative to the LoRA base directory.
recursive	BOOLEAN	false	Walk nested subfolders (relative paths).

Outputs: lora_names (newline-joined), total_count.

---

File tools

These read or write LoRA files on disk. The save nodes always write a new file to the ComfyUI output directory with an auto-incrementing counter — existing files are never overwritten and the source LoRA is never modified.

LoRA Scale & Save

Bakes a strength change into a LoRA file — by scaling its alphas, setting them outright, or scaling the weights — and saves the result as a new safetensors file.

A LoRA's contribution is (alpha / rank) · (up @ down), so its effective strength can be pre-baked either through the alpha scalars or the weight tensors. Use this to permanently tone a LoRA up or down, or to add alpha values to a LoRA that lacks them.

Parameter	Type	Default	Description
lora_name	choice	—	Source LoRA from the loras folder.
filename_prefix	STRING	loras/CCN_scaled_lora	Output path + prefix, relative to the output directory.
mode	choice	scale_alpha	See modes below.
value	FLOAT	1.0	Meaning depends on mode (-10–10).

Modes:

• scale_alpha — multiply every layer's alpha by value. Layers with no alpha get one created as rank × value, so this mode can add alphas to a LoRA that had none.
• set_alpha — write value as the exact alpha for every layer, replacing existing alphas.
• scale_weights — multiply the lora_up weights (and any diff/diff_b tensors) by value, leaving alphas alone.

Outputs: filepath (the absolute path of the new file).

fp8 tensors are upcast for the multiply and cast back, preserving their dtype. Training metadata is not carried over to the new file.

LoRA Truncate Rank

Reduces a LoRA's rank by slicing off the least-significant components — fast (seconds, no SVD re-decomposition) and produces a proportionally smaller file.

SVD-extracted LoRAs only

Slicing is only mathematically valid for SVD-extracted LoRAs (e.g. from a checkpoint-diff extract), where components are stored in descending order of importance. Trained LoRAs don't order their rank dimensions, so truncating one discards arbitrary components and degrades quality unpredictably. The node does not check this for you.

Alphas are rescaled automatically (alpha × new_rank / old_rank) so the LoRA's effective strength is preserved. Layers already at or below the target rank pass through unchanged. Training metadata is preserved and annotated with the truncation.

Parameter	Type	Default	Description
lora_name	choice	—	Source LoRA (kohya, PEFT, and ControlLoRA key formats supported).
new_rank	INT	32	Target rank (1–4096); per-layer the kept rank is min(new_rank, old_rank).
filename_prefix	STRING	loras/CCN_truncated_lora	Output path + prefix relative to the output directory.
output_dtype	choice	match_original	match_original, fp16, bf16, fp32, fp8_e4m3, fp8_e5m2. fp8 roughly halves size vs fp16.
verbose	BOOLEAN	true	Print per-layer rank changes and a size summary.

Outputs: filepath. The filename encodes the change, e.g. CCN_truncated_lora_trunc128to32_fp8_e4m3_00001.safetensors, and the console reports the before/after file size.

LoRA Metadata

Reads a LoRA file's embedded training metadata and surfaces it in a human-readable report — trigger words, base model, rank/alpha, learning rates, tag frequencies, and a detected architecture line.

It understands the metadata conventions of many trainers (kohya ss_*, modelspec, EveryDream2, ai-toolkit/Ostris, SimpleTuner, diffusers) and highlights trigger words gathered from several locations, including dataset folder names and the most frequent training tag. Use it to figure out how to actually use a downloaded LoRA. Read-only — nothing is modified.

Parameter	Type	Default	Description
lora_name	choice	—	LoRA file to inspect.
debug_mode	BOOLEAN	false	Also print the summary to the console.

Outputs: summary (the curated, sectioned report), full_metadata (a raw alphabetical dump of every metadata key).

Safetensors Metadata

A generic safetensors inspector for any model type — checkpoint, VAE, LoRA, CLIP, ControlNet, UNet, upscaler — reporting metadata plus the tensor structure.

Where LoRA Metadata curates training metadata for LoRAs specifically, this node works on any safetensors file and adds a structural overview: tensor count, dtype histogram, top-level module grouping, and individual tensor shapes. It runs the same architecture fingerprinting (detecting SD1.5/SDXL/Flux/SD3/Wan/HunyuanVideo/LTX families, LoRA formats, VAE types, text encoders) — handy for identifying an unknown file.

Parameter	Type	Default	Description
model_type	choice	—	Which ComfyUI model folder to look in (loras, checkpoints, vae, clip, unet, controlnet, …).
filename	STRING	""	The file. Exact match, unique case-insensitive substring match, or an absolute path all work.
show_tensors	BOOLEAN	true	Include the tensor-structure section.
max_tensors	INT	50	Cap on individually listed tensors.
debug_mode	BOOLEAN	false	Also print the summary to the console.

Outputs: summary, full_metadata.