ComfyUI Nifty Nodes

A ComfyUI node pack focused on workflow organisation and flexibility. The centrepiece is a Bundle system for passing arbitrary named values through a single wire, alongside smart media loaders with built-in resize, a multi-LoRA loader, a full set of logic and switch nodes, image processing tools, latent utilities, and various quality-of-life helpers.

Installation

Clone the repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/Stibo/comfyui-nifty-nodes

Then restart ComfyUI. Optional dependencies for certain nodes:

pip install kornia        # reinhard_lab_gpu and perceptual_crossfade in Image Color Match
pip install color-matcher # additional color transfer methods in Image Color Match

Features

• ComfyUI v3 native — built entirely on the comfy_api.latest API with proper define_schema, lazy evaluation, fingerprint_inputs caching and async execution where applicable
• Bundle system — pass any number of named values through a single BUNDLE wire, with dot-notation support for nested structures
• Smart media loaders — load images, videos or both with an integrated resize pipeline, optional frame rate control and embedded meta bundle extraction; use Embed Media Meta Bundle to inject metadata into saved PNGs
• Multi-LoRA loader — manage an entire LoRA stack in one node with per-row enable/disable and strength controls; or build reusable stacks with Lora Stack and Apply nodes
• Full logic suite — lazy and eager input switches, output routers, indexed combo switches, boolean operators, string/number comparisons and more
• Image Color Match — 10 color transfer methods including GPU-accelerated options, batch multithreading and selective channel transfer
• Model guidance — WAN-native NAG (Normalized Attention Guidance) and SLG/STG nodes for WAN 2.1 / 2.2 video models
• CLIP Text Encode — combined positive/negative prompt encoding in a single node
• Dynamic Prompt — wildcard-based prompt generation with deterministic seeds, powered by the dynamicprompts library
• Search aliases on every node for fast discovery via the node search

Note

• Tested with ComfyUI frontend 1.39.
• Untested with Nodes 2.0.
• The promotion of the special widgets to the subgraph node does not work correctly in version 1.39+.

---

Nodes

Bundle

The Bundle system lets you pack multiple values of any type into a single BUNDLE wire and unpack or access them wherever needed. This keeps graphs clean when many values need to travel together.

Bundle Pack

Packs up to 32 values of any type into a Bundle. An optional incoming Bundle can be merged — new values are added to it or overwrite existing keys.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	No	—	Existing bundle to merge into. New slot values overwrite keys with the same name.
value1 … value32	ANY	No	—	Values to pack. Controlled by count.
count	INT	No	1	Number of value slots to show (1–32).
hide_links	BOOL	No	False	Hide the connection wires going into this node to reduce visual clutter.

Outputs

Name	Type	Description
bundle	BUNDLE	The resulting bundle containing all packed values.

Bundle Unpack

Unpacks a Bundle back into up to 32 individual outputs. Output labels are set by renaming the slots, matching the keys used when packing.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	Yes	—	The bundle to unpack.
count	INT	No	1	Number of output slots to show (1–32).
hide_links	BOOL	No	False	Hide the connection wires going out of this node to reduce visual clutter.

Outputs

Name	Type	Description
bundle	BUNDLE	The original bundle, passed through unchanged.
value1 … value32	ANY	Unpacked values. Slots beyond count output None.

Bundle Get

Reads a single value from a Bundle by key.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	Yes	—	The bundle to read from.
key	STRING	Yes	value1	Key to read. Supports dot notation for nested keys, e.g. settings.width.

Outputs

Name	Type	Description
value	ANY	The value at the given key, or None if the key does not exist.

Bundle Set

Writes or overwrites a single key in a Bundle.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	Yes	—	The bundle to update.
key	STRING	Yes	value1	Key to write. Supports dot notation for nested keys, e.g. settings.width.
value	ANY	Yes	—	Value to store at the given key.

Outputs

Name	Type	Description
bundle	BUNDLE	Updated bundle. The original is not modified — a new copy is returned.

Slot Labels and Key Names

When you rename a slot in Bundle Pack or Bundle Unpack (double-click the slot label), the new name becomes the key used inside the bundle. The same label must be used on both ends for values to align correctly.

Dot Notation and Nesting

Keys use . as a path separator, allowing you to read and write into nested structures. For example:

Bundle Pack  →  slot label: "camera.fov"   →  value: 90
Bundle Get   →  key: "camera.fov"          →  returns: 90

You can nest arbitrarily deep: "render.output.format", "lora.0.strength", etc. Bundle Get and Bundle Set both support full dot-path access.

Merging Bundles

Bundle Pack accepts an optional incoming Bundle. Any values you pack are merged on top of it — existing keys are overwritten, other keys are preserved. This makes it easy to incrementally build up a shared parameter object across multiple nodes.

---

Media

Load & Resize Media

The most versatile loader — accepts images and videos in a single node, with a built-in resize pipeline. A is_video boolean output lets you route the result differently depending on what was loaded.

Inputs

Name	Type	Required	Default	Description
file	COMBO	Yes	none	File to load. Supports image formats (jpg, png, webp, …), animated formats (gif, apng, webp, avif) and video formats (mp4, mkv, webm, …).
force_frame_rate	INT	No	0	Enforce a specific frame rate. 0 = keep original. Values like 24, 30 or 60 resample the video during loading.
resize	DYNAMIC	No	off	Resize mode to apply after loading — see Resize Modes.

Outputs

Name	Type	Description
IMAGE	IMAGE	Loaded frames as a batch tensor. None if file is none.
frames	INT	Total number of frames (1 for still images).
width	INT	Output width after any resize.
height	INT	Output height after any resize.
frame_rate	INT	Frame rate of the source (or the forced rate if set).
audio	AUDIO	Extracted audio track. Silent if the file has no audio.
is_video	BOOL	True if the loaded file is a video or animation, False for still images.
meta_bundle	BUNDLE	Meta bundle embedded in the file's PNG metadata, if any. Empty dict if none found.

Load & Resize Video

Same as Load & Resize Media but restricted to video and animation file types.

Inputs / Outputs

Same as Load & Resize Media, minus the is_video output.

Load & Resize Image

Same as Load & Resize Media but restricted to image and animation file types. No frames, frame_rate or audio outputs.

Outputs

Name	Type	Description
IMAGE	IMAGE	Loaded image as a tensor. None if file is none.
width	INT	Output width after any resize.
height	INT	Output height after any resize.
meta_bundle	BUNDLE	Meta bundle embedded in the file's PNG metadata, if any. Empty dict if none found.

Embed Media Meta Bundle

Injects a BUNDLE dict into the workflow's extra_pnginfo so it gets embedded into any PNG saved downstream. Place this node in the wire between a loader and a Save Image node. Passes the input value through unchanged.

Inputs

Name	Type	Required	Default	Description
input	ANY	Yes	—	Any value to pass through unchanged.
meta_bundle	BUNDLE	Yes	—	The bundle to embed into the saved PNG metadata.
mode	COMBO	No	overwrite	overwrite = replace any existing meta_bundle entirely. merge = combine with an existing bundle — new keys override old ones, keys not present in the new bundle are kept.

Outputs

Name	Type	Description
output	ANY	The input value, passed through unchanged.

Resize Modes

All three loaders share the same resize pipeline, available via the resize dropdown:

Mode	Description
off	No resize — load at original dimensions.
scale dimensions	Scale to exact width × height. Optional center crop to avoid stretching.
scale by multiplier	Scale both dimensions by a float multiplier.
scale longer dimension	Scale so the longer side equals the given size, preserving aspect ratio.
scale shorter dimension	Scale so the shorter side equals the given size, preserving aspect ratio.
scale width	Scale to a specific width, preserving aspect ratio.
scale height	Scale to a specific height, preserving aspect ratio.
scale total pixels	Scale to a target megapixel count.
make divisible by	Round dimensions to the nearest multiple (e.g. 16 for WAN).

All modes include a divisible_by option and separate scale method selectors for upscaling and downscaling:

Scale Method	Best for
nearest-exact	Fastest, pixel art / sharp pixel edges
bilinear / bicubic	General purpose
area	Downscaling — sharp, minimal aliasing
lanczos	Upscaling — smooth, good detail retention
mitchell-netravali	Sharp upscaling with minimal ringing
bilinear (antialias) / bicubic (antialias)	Smooth downscaling with anti-aliasing
nvidia-rtx-vsr	AI upscaling — requires an NVIDIA RTX GPU and the nvidia-vfx package

Nifty Upscale Image

Loads an upscale model and upscales an image with it in a single node. Merges the core UpscaleModelLoader and ImageUpscaleWithModel nodes and adds tile parameters. On out-of-memory the tile size is halved automatically.

Inputs

Name	Type	Required	Default	Description
image	IMAGE	Yes	—	Image to upscale.
model	COMBO	Yes	—	Upscale model from the upscale_models folder.
tile_size	INT	No	512	Tile size used for tiled upscaling (64–4096).
overlap	INT	No	32	Overlap between tiles (0–4096).

Outputs

Name	Type	Description
IMAGE	IMAGE	Upscaled image.
model_name	STRING	Name of the used model without file extension.

Nifty Save Image

Saves images as PNG to a freely chosen folder. Optionally embeds the workflow metadata and a meta_bundle.

Inputs

Name	Type	Required	Default	Description
images	IMAGE	Yes	—	The images to save.
meta_bundle	BUNDLE	No	—	Embedded under the key meta_bundle together with the prompt and workflow metadata (only if save_metadata is enabled).
filename	STRING	No	ComfyUI	Filename prefix. Supports date placeholders like %date:yyyy-MM-dd% and %width%, %height%, %year%, %month%, %day%, %hour%, %minute%, %second%, %batch_num%.
filename_suffix	STRING	No	``	Appended to filename before the counter, e.g. _upscaled.
folder	STRING	No	output	Absolute path or relative to the ComfyUI root (e.g. ../output). Folders outside the registered output directory show no preview thumbnail, but the file is still saved.
save_metadata	BOOL	No	True	Embed prompt, workflow and meta_bundle as PNG text chunks.
overwrite	BOOL	No	False	False = a counter (_00001, _00002, …) is appended so existing files are never replaced. True = no counter, an existing file with the same name is overwritten. Ignored for batches with more than one image.

Outputs

Name	Type	Description
images	IMAGE	The input images, passed through unchanged.
filename	STRING	Comma-separated list of all saved filenames.

Batch Image Loader Start

Loads one image per loop iteration from a directory. Connect loop to Batch Image Loader End; everything between the two nodes runs once per image.

Inputs

Name	Type	Required	Default	Description
directory	STRING	No	output	Absolute path or relative to the ComfyUI root.
limit	INT	No	-1	Number of images to process. -1 = all, 0 = none.
start_index	INT	No	0	Number of images to skip at the start of the sorted list.
orderby	COMBO	No	name	Sort criterion: name (natural order: 1, 2, 10, 100), date modified, date created, size or type.
order	COMBO	No	ascending	ascending or descending.
recursive	BOOL	No	False	Also read images from subdirectories.
value1 … value16	ANY	No	—	Auto-growing list of initial values. From the second iteration on, the value of the same slot on Batch Image Loader End is used instead.

Outputs

Name	Type	Description
loop	NIFTY_BATCH_LOOP	Connect to Batch Image Loader End.
image	IMAGE	Image of the current iteration.
filename	STRING	Path relative to the directory, including subfolders.
meta_bundle	BUNDLE	Meta bundle embedded in the PNG metadata of the current image, if any.
value1 … value16	ANY	Value of the current iteration. Pass it on to the same slot of Batch Image Loader End.

Batch Image Loader End

Applies a file action and a rename action to the image of the current iteration, then starts the next iteration or finishes the loop.

Inputs

Name	Type	Required	Default	Description
loop	NIFTY_BATCH_LOOP	Yes	—	Connect to Batch Image Loader Start.
file_action	DYNAMIC	Yes	none	none = keep the file · move = move it to target_directory (absolute or relative to the loader's directory, default done, created if missing, subfolders are preserved) · delete = delete the file. Runs before rename_action.
rename_action	DYNAMIC	No	none	Applied after the file action (in the new location if moved), not available when the file is deleted. none · rename = filename prefix plus counter (name_00001_.ext) · suffix = appends suffix to the file name · regex = applies regex_search / regex_replace to the file name without extension.
value1 … value16	ANY	No	—	Auto-growing list of values passed to the same slot of Batch Image Loader Start in the next iteration. Unconnected slots keep their current value.

Outputs

Name	Type	Description
count	INT	Number of processed images.
filenames	STRING	All processed file names, one per line. Only available after the last iteration.
value1 … value16	ANY	Value of this slot after the last iteration.

Crop Image

Crops one or more images without scaling. Width and height can be cropped independently from the left, center or right / top, center or bottom.

Inputs

Name	Type	Required	Default	Description
image	IMAGE	Yes	—	Image or image batch to crop.
width	INT	No	512	Target crop width in pixels. 0 = keep original width.
height	INT	No	512	Target crop height in pixels. 0 = keep original height.
crop_width	COMBO	No	center	Horizontal crop position: left, center or right.
crop_height	COMBO	No	center	Vertical crop position: top, center or bottom.

Outputs

Name	Type	Description
IMAGE	IMAGE	Cropped image.

---

Lora

Nifty Lora Loader

Applies a full LoRA stack to a model and CLIP in one node. Each row has its own enable toggle and strength slider. Rows with strength 0.0 or disabled are skipped entirely.

Inputs

Name	Type	Required	Default	Description
model	MODEL	Yes	—	Base model to apply LoRAs to.
clip	CLIP	No	—	CLIP model to apply LoRAs to.
loras	NIFTY_LORA_LOADER	Yes	—	LoRA stack widget. Add rows, pick a LoRA file, set strength and toggle enabled. Strength 1.0 = full effect, 0.5 = half, negative values invert.

Outputs

Name	Type	Description
MODEL	MODEL	Model with all enabled LoRAs applied in order.
CLIP	CLIP	CLIP with all enabled LoRAs applied in order.

Nifty Lora Stack

Builds a reusable NIFTY_LORA_STACK wire that can be passed to Nifty Apply Lora Stack. Accepts an optional incoming stack to chain multiple Lora Stack nodes together — new rows are appended. Rows with strength 0.0 or disabled are skipped.

Inputs

Name	Type	Required	Default	Description
lora_stack	NIFTY_LORA_STACK	No	—	Existing stack to extend. New rows are appended after it.
loras	NIFTY_LORA_LOADER	Yes	—	LoRA stack widget. Add rows, pick a LoRA file, set strength and toggle enabled. Strength 1.0 = full effect, 0.5 = half, negative values invert.

Outputs

Name	Type	Description
lora_stack	NIFTY_LORA_STACK	Combined LoRA stack ready to be passed to Nifty Apply Lora Stack.

Nifty Apply Lora Stack

Applies a NIFTY_LORA_STACK to a model and optional CLIP. Separates stack definition from application — useful when the same stack needs to be applied to multiple models or at different points in a workflow.

Inputs

Name	Type	Required	Default	Description
lora_stack	NIFTY_LORA_STACK	Yes	—	The LoRA stack to apply.
model	MODEL	Yes	—	Base model to apply LoRAs to.
clip	CLIP	No	—	CLIP model to apply LoRAs to.

Outputs

Name	Type	Description
MODEL	MODEL	Model with all LoRAs from the stack applied in order.
CLIP	CLIP	CLIP with all LoRAs from the stack applied in order.

---

Conditioning

Nifty CLIP Text Encode (Prompt)

Encodes positive and negative text prompts using a CLIP model in a single node. Both outputs are CONDITIONING wires ready for a sampler.

Inputs

Name	Type	Required	Default	Description
clip	CLIP	Yes	—	The CLIP model used for encoding the text.
positive	STRING	No	``	Positive text prompt. Leave empty for an unconditional embedding.
negative	STRING	No	``	Negative text prompt. Leave empty for an unconditional embedding.

Outputs

Name	Type	Description
positive	CONDITIONING	Positive conditioning from the encoded prompt.
negative	CONDITIONING	Negative conditioning from the encoded prompt.

---

Model

WAN-specific model patching nodes. Both are marked experimental and live in the nifty/model category.

Wan Video NAG

Applies Normalized Attention Guidance (NAG, NeurIPS 2025) to WAN 2.1 / 2.2 video models (T2V and I2V). Replaces or complements CFG by applying guidance directly in attention space — stable at low step counts. Forked and significantly reworked from KJNodes.

Inputs

Name	Type	Required	Default	Description
enable	BOOL	No	True	When False, model and conditioning pass through unchanged.
model	MODEL	Yes	—	WAN model to patch.
conditioning	CONDITIONING	Yes	—	Negative conditioning (empty or negative prompt). NAG uses this as its internal reference — the direction guidance is pushed away from in attention space. Do not connect positive conditioning here.
nag_scale	FLOAT	No	11.0	Guidance strength (φ), analogous to CFG scale. Good range: 8–15. 0 disables NAG. Tune last — settle on nag_tau and nag_alpha first.
nag_alpha	FLOAT	No	0.25	Blend weight (α) between NAG-guided and original positive attention. 0 = no effect, 1 = full NAG. Lower for I2V (0.10–0.20) to preserve the reference image.
nag_tau	FLOAT	No	2.5	L1-norm clipping threshold (τ): caps how far guided attention deviates from positive attention. Lower = safer. Higher = stronger. Lower for I2V (1.5–2.0); higher for few-step Lightning LoRA (3–5).
nag_sigma_end	FLOAT	No	0.0	NAG is skipped when sigma drops below this value. 0.0 = always active (correct for two-model HN/LN pipelines). For single-model pipelines, 0.75 gives near-identical quality with less compute.
nag_scale_end	FLOAT	No	0.0	Adaptive scale end: linearly interpolates from nag_scale (sigma=1.0) to this value (sigma=0.0). 0.0 = disabled (constant scale).
input_type	COMBO	No	default	default = sampler sends a [positive, negative] batch pair (standard CFG). batch = single conditioning without a paired negative (use for CFG=1 / few-step WAN with Lightning LoRA).
inplace	BOOL	No	False	Modify tensors in-place to reduce peak VRAM. Slightly alters numerical results — enable only if out of memory.
cond_zero_out	BOOL	No	True	Output a zeroed neutral conditioning instead of the input conditioning. Recommended — NAG handles guidance internally so the sampler's negative slot should be empty to avoid interference. Disable only if intentionally stacking NAG on top of CFG.

Outputs

Name	Type	Description
MODEL	MODEL	Patched model with NAG applied to every cross-attention block.
CONDITIONING	CONDITIONING	The input conditioning, zeroed out if cond_zero_out is True.

Recommended Starting Values

Scenario	nag_scale	nag_tau	nag_alpha
T2V multi-step	11	2.5	0.25
T2V few-step (Lightning LoRA)	11–15	2.5–5	0.25–0.5
I2V (any)	11	1.5–2.0	0.10–0.20

Nifty Wan Video SLG

Applies Spatiotemporal Skip Guidance (STG / SLG) to WAN video models. No TeaCache required. Works at CFG=1 (e.g. LightX2V). Paper: arxiv 2411.18664 (CVPR 2025). Forked and significantly reworked from KJNodes.

Inputs

Name	Type	Required	Default	Description
enabled	BOOL	No	True	When False, the model is passed through unchanged.
model	MODEL	Yes	—	WAN model to patch.
blocks	STRING	No	9, 10	Comma-separated block indices. Supports ranges (9-11 = 9, 10, 11). WAN 2.2 has 40 layers. Blocks 9, 10 are empirically recommended. Start with 10 alone if flickering occurs.
mode	COMBO	No	STG-A	STG-A — zeros self-attention only (cross-attn and FFN still run). Less perturbation, lower flickering risk. Recommended for WAN + few-step. STG-R — entire block becomes identity. Stronger but more prone to flickering at 2–3 steps.
scale	FLOAT	No	1.0	Guidance strength. Works at CFG=1. LightX2V (2–3 HN steps): STG-A max ~2.0–2.5, STG-R max ~1.5.
start_percent	FLOAT	No	0.0	Start of active range as a fraction of the full denoising schedule (HN + LN combined). Percentages are absolute.
end_percent	FLOAT	No	1.0	End of active range. Formula: HN_steps / total_steps. 4-step (2+2): 0.5. Default 1.0 = active for all sigmas this model processes.
rescaling_scale	FLOAT	No	0.0	Normalises correction std to match cond_pred std before applying scale. 0.0 = disabled. 0.3–0.5 recommended if scale causes flickering.
scale_end	FLOAT	No	0.0	Adaptive scale end: interpolates from scale at start_percent to scale_end at end_percent. 0.0 = disabled (constant scale).

Outputs

Name	Type	Description
MODEL	MODEL	Model with a post_cfg_function patched in to apply STG on each sampler step.

---

Logic

A full set of flow control nodes. Most come in two variants: lazy (only evaluates the branch that will actually be used) and eager (evaluates both branches regardless).

Input Switch

Routes one of two inputs to the output based on a boolean. Lazy — the unused branch is never executed.

Inputs

Name	Type	Required	Default	Description
on_true	ANY	No	—	Passed through when boolean is True.
on_false	ANY	No	—	Passed through when boolean is False.
boolean	BOOL	Yes	—	Selector.

Outputs

Name	Type	Description
output	ANY	The selected input value.

Input Switch (Eager)

Same as Input Switch but both branches are always evaluated.

None Input Switch

Routes between on_exists and on_none depending on whether value is None. Lazy.

Inputs

Name	Type	Required	Default	Description
on_exists	ANY	No	—	Passed through when value is not None.
on_none	ANY	No	—	Passed through when value is None.
value	ANY	No	—	The value to check.

Outputs

Name	Type	Description
output	ANY	The selected branch.

None Input Switch (Eager)

Same as None Input Switch but both branches are always evaluated.

Index Input Switch

Routes one of up to 16 inputs to the output by index. Lazy — only the selected input is evaluated.

Inputs

Name	Type	Required	Default	Description
index	INT	Yes	1	Which input to pass through (1–16).
value1 … value16	ANY	No	—	Auto-growing input list.

Outputs

Name	Type	Description
output	ANY	The value at the selected index, or None if the slot is not connected.

Index Input Switch (Eager)

Same as Index Input Switch but all inputs are always evaluated.

Output Switch

Routes a single input to either on_true or on_false. The unselected output is blocked (downstream nodes do not execute).

Inputs

Name	Type	Required	Default	Description
input	ANY	Yes	—	Value to route.
boolean	BOOL	No	True	True = route to on_true, False = route to on_false.

Outputs

Name	Type	Description
on_true	ANY	Active when boolean is True, blocked otherwise.
on_false	ANY	Active when boolean is False, blocked otherwise.

None Output Switch

Routes input to either on_exists or on_none depending on whether value is None.

Inputs

Name	Type	Required	Default	Description
input	ANY	No	—	Value to route.
value	ANY	No	—	Value to check for None.

Outputs

Name	Type	Description
on_exists	ANY	Active when value is not None.
on_none	ANY	Active when value is None.

Index Output Switch

Routes a single input to one of up to 16 outputs by index. All other outputs are blocked.

Inputs

Name	Type	Required	Default	Description
input	ANY	No	—	Value to route.
index	INT	Yes	1	Which output to route to (1–16).

Outputs

Name	Type	Description
value1 … value16	ANY	Only the output at index is active; all others are blocked.

Signal Switch

Simple on/off gate. When passthrough is False, the output is blocked and all downstream nodes skip execution.

Inputs

Name	Type	Required	Default	Description
input	ANY	Yes	—	Value to gate.
passthrough	BOOL	No	True	True = pass through, False = block.

Outputs

Name	Type	Description
output	ANY	The input value, or blocked if passthrough is False.

First Switch

Returns the first non-None value from a list of up to 16 inputs. Lazy — inputs are evaluated one by one until a non-None value is found.

Inputs

Name	Type	Required	Default	Description
value1 … value16	ANY	No	—	Auto-growing input list checked in order from top to bottom.

Outputs

Name	Type	Description
output	ANY	First non-None value found, or None if all inputs are None.

String Compare

Compares a string against a match value and returns a boolean.

Inputs

Name	Type	Required	Default	Description
string	STRING	Yes	—	The string to test.
match	STRING	Yes	``	The value to compare against. An empty match always returns False.
comparison	DYNAMIC	Yes	exact	exact = full equality · contains = substring · starts with / ends with = prefix/suffix · regex = regular expression
case_sensitive	BOOL	No	False	Enable case-sensitive comparison.
negate	BOOL	No	False	Invert the result.

Outputs

Name	Type	Description
BOOL	BOOL	Result of the comparison.

Number Compare

Compares two numbers and returns a boolean.

Inputs

Name	Type	Required	Default	Description
a	INT / FLOAT	Yes	—	Left-hand operand.
b	INT / FLOAT	Yes	—	Right-hand operand.
comparison	DYNAMIC	Yes	a == b	a == b · a != b · a < b · a > b · a <= b · a >= b · a <= b <= c (range check)
negate	BOOL	No	False	Invert the result.

Outputs

Name	Type	Description
BOOL	BOOL	Result of the comparison.

Is None

Returns True if the input is None (not connected or explicitly None).

Inputs

Name	Type	Required	Default	Description
input	ANY	No	—	Value to check.
negate	BOOL	No	False	Invert — returns False when None, True otherwise.

Outputs

Name	Type	Description
BOOL	BOOL	True if input is None, False otherwise (before negation).

Int Switch / Float Switch / String Switch

Each returns one of two primitive values based on a boolean.

Inputs

Name	Type	Required	Default	Description
boolean	BOOL	No	True	Selector.
on_true	INT / FLOAT / STRING	No	0 / 0.0 / ``	Returned when boolean is True.
on_false	INT / FLOAT / STRING	No	0 / 0.0 / ``	Returned when boolean is False.

Combo Switch

Switches between two combo (dropdown) values based on a boolean. Each side has its own options list, defined as a pipe-separated string (option1|option2|option3). Also outputs the selected value as a STRING and the index within its option list.

Inputs

Name	Type	Required	Default	Description
boolean	BOOL	No	True	Selector.
on_true	COMBO	No	—	Selected value when True.
true_options	STRING	No	—	Available options for on_true, separated by |.
on_false	COMBO	No	—	Selected value when False.
false_options	STRING	No	—	Available options for on_false, separated by |.

Outputs

Name	Type	Description
COMBO	COMBO	Selected combo value.
string	STRING	The selected combo value as a string.
index	INT	Index of the selected value within its option list. -1 if not found.

Index Combo Switch

Like Index Input Switch but selects the active branch via a named dropdown instead of a raw integer. Each branch gets a custom label defined by the option1–option16 string inputs — the dropdown (choise) is built from those labels at runtime. Lazy — only the selected branch is evaluated.

Inputs

Name	Type	Required	Default	Description
values	ANY	No	—	Auto-growing input list. Only the selected branch is evaluated (lazy).
choise	COMBO	Yes	—	Dropdown showing the labels defined by option1–option16.
option1 … option16	STRING	No	Option N	Label for each corresponding value input. Defines the available dropdown options.

Outputs

Name	Type	Description
output	ANY	Value of the selected branch, or None if not connected.
COMBO	COMBO	The selected label as a COMBO value.
string	STRING	The selected label as a string.
index	INT	0-based index of the selected option. -1 if not found.

Index Combo Switch (Eager)

Same as Index Combo Switch but all inputs are always evaluated regardless of which option is selected.

Boolean AND / OR / XOR

Standard two-input boolean gates.

Boolean AND All / OR Any

Multi-input variants using an auto-growing input list (up to 16). AND All returns True only if every connected input is True. OR Any returns True if at least one is True.

Boolean NOT

Inverts a boolean value.

None

Outputs a None value. Useful as an explicit empty input for optional slots.

---

Image

Image Color Match

Transfers the color distribution of a reference image onto a target image. Supports 10 transfer methods, selective channel transfer, adjustable strength and batch multithreading.

Inputs

Name	Type	Required	Default	Description
image_target	IMAGE	No	—	Image whose colors will be adjusted.
image_ref	IMAGE	No	—	Reference image providing the color distribution.
enabled	BOOL	No	True	When disabled, image_target passes through unchanged.
method	COMBO	No	mkl	Color transfer algorithm — see Methods below.
transfer_mode	COMBO	No	all	all = full color transfer · color = hue/chroma only, keep original luminance · luminance = brightness only, keep original colors
strength	FLOAT	No	1.0	Blend amount. 1.0 = full match, 0.5 = half-way blend, values above 1.0 exaggerate the effect. Range 0.0–10.0.
multithread	BOOL	No	True	Process batch frames in parallel using multiple CPU threads. Recommended for large batches.

Outputs

Name	Type	Description
IMAGE	IMAGE	Color-matched image batch. None if either input is not connected.

Color Match Methods

Method	Notes
mkl	Monge-Kantorovich linear map — best overall quality, recommended default for most use cases
reinhard_lab	Fast statistical LAB match — good for natural/photographic content where speed matters. CPU only
reinhard_lab_gpu	Same quality as reinhard_lab but GPU-accelerated — prefer this over reinhard_lab when kornia is available. Requires kornia
wavelet	Transfers low-frequency color while preserving structure — best when texture and fine detail must stay intact
adain	Adaptive instance normalisation — fast and consistent, well suited for video/animation frame batches
hm	Histogram matching — fast broad tonal correction, less accurate for complex or saturated palettes
reinhard	Classic Reinhard RGB — quick and simple, struggles with complex colour distributions
mvgd	Multivariate Gaussian — high-accuracy statistical match, good complement to mkl for tricky scenes. Requires color-matcher
hm-mvgd-hm	Histogram + MVGD combo — robust for challenging distributions when mkl alone falls short. Requires color-matcher
hm-mkl-hm	Histogram + MKL combo — highest overall quality among the combo methods. Requires color-matcher

Merge Image Batches

Concatenates two image batches with optional blending at the transition. Useful for stitching video segments.

Inputs

Name	Type	Required	Default	Description
source_images	IMAGE	No	—	First batch.
new_images	IMAGE	No	—	Second batch, appended after source.
overlap_mode	DYNAMIC	No	none	Transition mode — see below.

Overlap Mode	Description
none	Simple concatenation, no blending.
cut	Remove overlap frames from the selected side before concatenating.
linear_blend	Simple linear crossfade over the overlap region.
ease_in_out	S-curve (smoothstep) crossfade — more natural motion.
filmic_crossfade	Gamma-correct (2.2) crossfade — preserves perceived brightness.
perceptual_crossfade	LAB color space crossfade — perceptually even transition. Requires kornia.

Outputs

Name	Type	Description
images	IMAGE	Merged batch. None if both inputs are disconnected.

Resize Image

Resizes an image or mask using the full Nifty resize pipeline. Accepts both IMAGE and MASK types.

Inputs

Name	Type	Required	Default	Description
input	IMAGE / MASK	Yes	—	Image or mask to resize.
resize	DYNAMIC	Yes	—	Resize mode — see Resize Modes.

Outputs

Name	Type	Description
resized	IMAGE / MASK	Resized output (same type as input).
width	INT	Output width.
height	INT	Output height.

Image From Batch

Extracts a slice from an image batch by start index and length.

Inputs

Name	Type	Required	Default	Description
image	IMAGE	Yes	—	Input batch.
batch_index	INT	No	0	Start index. Positive = from beginning, negative = from end (-1 = last image). Range −4096–4096.
length	INT	No	1	Number of images to take. -1 = all from start to end of batch. Range −1–4096.

Outputs

Name	Type	Description
IMAGE	IMAGE	Extracted slice.

Last Image From Batch

Shortcut to take the last N images from a batch.

Inputs

Name	Type	Required	Default	Description
image	IMAGE	Yes	—	Input batch.
length	INT	No	1	Number of images to take from the end (1–4096).

Outputs

Name	Type	Description
IMAGE	IMAGE	Last length images from the batch.

---

Latent

VAE Encode

Encodes an image batch into latents. Automatically detects the temporal compression factor of the VAE and can optionally encode only a subset of frames specified by a target latent count.

Inputs

Name	Type	Required	Default	Description
pixels	IMAGE	Yes	—	Image frames to encode.
vae	VAE	Yes	—	VAE model to use for encoding.
target_latents	INT	No	0	0 = encode all frames. Positive = encode only the first N latents' worth of frames. Negative = encode only the last N latents' worth. The temporal factor is detected automatically from the VAE.

Outputs

Name	Type	Description
LATENT	LATENT	Encoded latent samples.
IMAGE	IMAGE	The original pixel frames, passed through unchanged.

Normalize Video Latent Start

Normalises the statistics (mean and standard deviation) of the first few latent frames to match the subsequent frames. Helps reduce flicker or discontinuity at the beginning of a generated video.

Inputs

Name	Type	Required	Default	Description
latent	LATENT	Yes	—	Video latent to process. Passthrough if the latent has only 1 frame.
enabled	BOOL	No	True	When disabled, the latent is passed through unchanged.
start_frame_count	INT	No	4	Number of latent frames from the start to normalise (1–max).
reference_frame_count	INT	No	5	Number of frames immediately after the start region to use as the normalisation reference (1–max).

Outputs

Name	Type	Description
LATENT	LATENT	Latent with normalised start frames.

Latent From Batch

Extracts a slice from a latent batch by start index and length. Supports both 4D (image) and 5D (video) latents.

Inputs

Name	Type	Required	Default	Description
samples	LATENT	Yes	—	Input latent batch.
batch_index	INT	No	0	Start index. Positive = from beginning, negative = from end (-1 = last frame). Range −4096–4096.
length	INT	No	0	0 = all from start to end. Positive = first N frames. Negative = trim N frames from the opposite end (e.g. -1 = all except the last).

Outputs

Name	Type	Description
LATENT	LATENT	Extracted latent slice.

Last Latent From Batch

Shortcut to take the last N latents from a batch.

Inputs

Name	Type	Required	Default	Description
samples	LATENT	Yes	—	Input latent batch.
length	INT	No	1	Number of latents to take from the end (1–4096).

Outputs

Name	Type	Description
LATENT	LATENT	Last length latents from the batch.

---

Sampling

Nifty Random Noise

Generates reproducible random noise from a seed for use with advanced sampler nodes that accept a NOISE input.

Inputs

Name	Type	Required	Default	Description
seed	INT	No	0	Seed value for noise generation (0–1125899906842624).

Outputs

Name	Type	Description
NOISE	NOISE	Noise object ready for use in an advanced KSampler node.

Flux 2 Sampler

Experimental all-in-one Flux.2 sampler. Combines FluxGuidance, BasicGuider, KSamplerSelect, Flux2Scheduler, Empty Flux 2 Latent and SamplerCustomAdvanced. Lives in the sampling/custom_sampling category.

Inputs

Name	Type	Required	Default	Description
model	MODEL	Yes	—	Model to sample with.
conditioning	CONDITIONING	Yes	—	Positive conditioning without embedded guidance. The guidance value is embedded internally (same as FluxGuidance).
latent_image	LATENT	No	—	Latent to denoise. If unconnected, an empty latent is generated from width / height.
sampler	SAMPLER	No	—	Sampler override. If unconnected, sampler_name is used.
noise_seed	INT	No	0	Seed for noise generation (0–1125899906842624).
steps	INT	No	20	Number of sampling steps (1–4096).
guidance	FLOAT	No	3.5	Flux guidance embed strength (0.0–100.0).
sampler_name	COMBO	No	—	Sampler used when no sampler input is connected.
width	INT	No	1024	Latent width. Used for the sigma schedule and the empty latent only if latent_image is unconnected.
height	INT	No	1024	Latent height. Used for the sigma schedule and the empty latent only if latent_image is unconnected.

Outputs

Name	Type	Description
latent	LATENT	Sampled latent.
time	INT	Duration of the sampling call only, in milliseconds.

---

Selectors

Lightweight selector nodes that output a filename or name string. They exist to keep your actual loader nodes clean — connect these to a loader and swap models or settings without touching the loader itself.

Diffusion Model Selector

Selects a diffusion model file from the unet / diffusion_models folder. Supports .safetensors, .gguf, .pt, .pth, .bin.

Outputs

Name	Type	Description
model	ANY	Selected model filename string.
is_gguf	BOOL	True if the selected file has a .gguf extension.

Clip Selector

Selects a CLIP / text encoder model file from the clip / text_encoders folder.

Outputs

Name	Type	Description
clip	ANY	Selected model filename string.
is_gguf	BOOL	True if the selected file has a .gguf extension.

CLIP Type Selector

Selects a CLIP type string from a dropdown. Covers all model families supported by ComfyUI's dual-CLIP loader (stable_diffusion, wan, flux2, sd3, and many more). Connect the output to a CLIP loader's type input.

Outputs

Name	Type	Description
type	ANY	The selected CLIP type as a string.

Sampler Selector

Selects a sampler name from the full list of available ComfyUI samplers.

Scheduler Selector

Selects a scheduler name from the full list of available ComfyUI schedulers.

Upscale Model Selector

Selects an upscale model file from the upscale_models folder without loading it.

Outputs

Name	Type	Description
model_name	ANY	Selected model filename string. None if none is selected.

---

String

Any To String

Converts one or more values of any type to a single string. Booleans become true/false, None becomes "none", everything else uses str().

Inputs

Name	Type	Required	Default	Description
value1 … value16	ANY	No	—	Auto-growing list of values to convert. None values are excluded.
mode	DYNAMIC	No	delimiter	delimiter = join all values with a separator string · pattern = insert values into a template using $1, $2, … placeholders (e.g. "$1_$2" with hello and world → "hello_world")

Outputs

Name	Type	Description
string	STRING	Resulting string.

Bundle To String

Combines the scalar values (string, int, float, bool) of a Bundle into a single string. Booleans become true/false.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	Yes	—	Bundle to convert.
mode	DYNAMIC	No	delimiter	delimiter = join all scalar values of the first level with a separator string (default -) · pattern = insert bundle values into a template using {key} placeholders (default {name}). Nested values are accessed with {key.subkey}. Missing keys and non-scalar values produce an empty string.

Outputs

Name	Type	Description
string	STRING	Resulting string.

String Split

Splits a string into a list by a delimiter.

Inputs

Name	Type	Required	Default	Description
string	STRING	Yes	—	String to split. Must be connected.
delimiter	STRING	No	\n	Split character or string. Use \n for newline, \t for tab.
trim	BOOL	No	False	Trim leading and trailing whitespace from each part.

Outputs

Name	Type	Description
list	STRING	List of string parts.

String Join

Joins a list of strings into a single string.

Inputs

Name	Type	Required	Default	Description
list	STRING	Yes	—	List to join. Must be connected.
delimiter	STRING	No	\n	Separator inserted between each item. Use \n for newline, \t for tab.

Outputs

Name	Type	Description
string	STRING	Joined string.

Nifty Dynamic Prompt

Generates a resolved prompt string from a template using the dynamicprompts library. Supports wildcard files, random choices, weighted variants, combinatorial expansion and more — see the dynamicprompts documentation for the full syntax.

Wildcards are text files containing one option per line. Place them in your ComfyUI user directory at user/default/wildcards/ (create the folder if it doesn't exist). Reference them in your template with double underscores: __style__ loads user/default/wildcards/style.txt. Subdirectories work too: __artists/painters__ loads user/default/wildcards/artists/painters.txt.

The seed input makes generation deterministic — the same template and seed always produce the same result. If dynamicprompts is not installed, the template is returned unchanged.

The prompt field acts as a manual override: if it contains text, it is passed through directly without running the generator. Clear it to re-generate from template.

See full syntax reference: https://pypi.org/project/dynamicprompts/

Inputs

Name	Type	Required	Default	Description
prompt	STRING	No	``	Manual prompt override (socketless). If non-empty, bypasses the template generator and passes this value through directly.
template	STRING	No	``	Dynamic prompt template (socketless). Used when prompt is empty. Supports wildcard syntax, {a|b|c} choices, weights, ranges and more.
enabled	BOOL	No	True	When False, prompt and template pass through unchanged without generation.
seed	INT	No	0	Deterministic seed. Same template + seed always produces the same output.
cleanup_prompt	BOOL	No	True	Cleans up the generated prompt: removes unnecessary whitespaces, commas, periods and empty brackets.

Outputs

Name	Type	Description
prompt	STRING	The resolved final prompt string.
template	STRING	The original template, passed through unchanged.
seed	INT	The seed value, passed through unchanged.

---

Number

Nifty Math

Evaluates a free-form mathematical expression using up to 16 named inputs (a–p). Unconnected slots default to 0. The result is cast to all three output types simultaneously.

Inputs

Name	Type	Required	Default	Description
a … p	INT / FLOAT	No	0	Auto-growing variable inputs. The first two (a, b) are always visible; additional slots appear as needed up to 16.
expression	STRING	Yes	a + b	Expression to evaluate. See supported syntax below.

Outputs

Name	Type	Description
int	INT	Result cast to integer (truncated, not rounded).
float	FLOAT	Result as a float.
boolean	BOOL	False if result is 0 or 0.0, True otherwise.

Supported Syntax

Category	Examples
Arithmetic	a + b, a - b, a * b, a / b, a // b, a % b, a ** b
Comparisons	a == b, a != b, a < b, a > b, a <= b, a >= b
Boolean logic	a and b, a or b, not a
Ternary	a if a > 0 else b
Math functions	abs, round, min, max, pow, sqrt, floor, ceil, log, log2, log10, exp
Trig	sin, cos, tan, asin, acos, atan, atan2, radians, degrees, hypot
Helpers	clamp(v, lo, hi), lerp(a, b, t), remap(v, a, b, c, d)
Constants	pi, tau, e, inf

Division by zero returns 0. Any other expression error raises a descriptive exception.

---

Nifty Seed

A seed node with integrated New Seed and Random on Prompt controls. Clicking New Seed immediately randomises the seed. Random on Prompt automatically generates a new random seed every time the queue is triggered.

Inputs

Name	Type	Required	Default	Description
seed	INT	No	0	Seed value (0–1125899906842624).

Outputs

Name	Type	Description
seed	INT	The seed value, clamped to the valid range.

---

Output List

Nodes that output a list. Downstream nodes run once per list item.

Bundle Grid List

Generates every combination (cartesian product) of the given fields as a list of Bundles, e.g. for grid testing. Optionally merged with a base Bundle.

Inputs

Name	Type	Required	Default	Description
bundle	BUNDLE	No	—	Optional base values merged into every generated combination.
values1 … valuesN	STRING	No	``	Auto-growing list of fields in the format name=value1,value2,value3. Supports dot notation for nested keys, e.g. settings.width=512,768. Numeric values are converted to INT/FLOAT. Empty fields are ignored.

Outputs

Name	Type	Description
list	BUNDLE	List with one Bundle per combination.

Combo Auto List

Experimental. Connect combo to any COMBO input to run the downstream node once per combo value, e.g. to sweep every sampler in a single queue run.

Inputs

Name	Type	Required	Default	Description
combo	COMBO	No	—	Shows all values of the connected combo widget. Display only.
exclude	STRING	No	``	Comma-separated values to skip, supports * wildcards (e.g. euler,*_cfg,*sde*).

Outputs

Name	Type	Description
combo	ANY	List of all combo values except the excluded ones.

Any To List

Combines connected values of any matching type into a single output list.

Inputs

Name	Type	Required	Default	Description
item1 … item32	ANY	No	—	Auto-growing list of values. Lists are flattened into the result.

Outputs

Name	Type	Description
list	ANY	List containing all connected values.

---

Utils

Experimental workflow organisation and debugging nodes.

Hidden Link

Passes any value through unchanged while hiding the connection wire on the incoming side. Useful for keeping graphs readable when a connection has to cross a large distance.

Bypass By Title

Bypasses a set of nodes by title when a boolean is True. Node titles are matched case-insensitively and support partial matches.

Inputs

Name	Type	Required	Default	Description
bypass	BOOL	No	False	When True, the listed nodes are bypassed.
nodes	STRING	No	``	Node titles to bypass, one per line.
search_from_root	BOOL	No	True	Search in the root graph even when called from inside a subgraph.
enforce	BOOL	No	True	Actively enforce the bypass state so it cannot be overridden.

Bypass Switch By Title

A multi-option version of Bypass By Title. Each option has its own label and list of nodes to bypass. Selecting an option bypasses all nodes in that option's list and activates the others.

Node Chain Extender

Dynamically duplicates a node and chains the copies together, connecting specified slots between adjacent links.

Subgraph Labels

Assigns display labels to up to 10 subgraph ports. Experimental — requires subgraph support in your ComfyUI version.

Simple Title

A display-only node showing a text label in the graph. Useful for annotating sections of a workflow.

Debug Any

Displays any value as a text preview below the node. Primitives are shown as-is; complex types are serialized to JSON.

Preview Any

Displays up to 16 values simultaneously as inline text previews. For image tensors, shows a summary (N images, WxH) rather than rendering the images. Output lists are shown as one preview per item. The previews are persistent and stay after reloading the workflow.