KJNodes for ComfyUI

Burn text labels onto your images

Roughing up a motion track so it doesn't look robotic

Merging two InstanceDiffusion tracks into one

Build a string list from two inputs

Push HunyuanVideo past its frame limit

Push Wan past its frame limit (sometimes)

Stitching two audio clips end to end

Text-prompt masks across a whole batch

Crop a whole frame batch to one consistent mask box

Crop a moving subject out of a video, jitter-free

Pasting your processed crops back where they came from

Pasting detail crops back where they came from

Unpack a bounding box into plain integers

Draw detection boxes onto your image

Snap a mask to a coarse block grid

One true/false switch to drive a whole workflow

Checking a camera path before you render it

The CFG-Zero* trick, wrapped as a model patch

Load a checkpoint with the speed hacks built in

Nudging a model's weights, block by block

Match one image's color grade to another

Match one image's color grade to another

Turn a specific RGB color into a mask

Merge many conditionings without a node pyramid

Regional prompting for two areas in one node

The same regional-prompting node, three regions deep

Four-region regional prompting in one node

The pack's biggest bundled regional-prompting node

A bypassable stand-in for your positive/negative pair

Packing a pile of small masks into fewer tiles

Actually see how your context windows tile a long video

Deprecated, and the only audio-reactive mask node in the pack

Deprecated three-point fade — check Cross Fade Images first

Keyframe a mask's value over a batch of frames

An animated mask from a tiny fluid sim

A gradient anchored to a moving point

A smooth gradient mask, still or animated

Per-object prompts via bounding boxes

Procedural blobby masks that animate over frames

Drawing a moving dot from coordinates

A circle, square, or triangle mask on demand (and animated)

Animated shape masks that move along a path

Rendering (and animating) text as a mask

Animating text along a set of coordinates

An animated cellular-pattern mask

Blending two clips together with a real transition

Turn a stack of images into a transitions video

A per-step weight schedule for Advanced-ControlNet

Pasting a hand-picked noise schedule straight in

The After Effects trick, but inside ComfyUI

Decode long video straight to disk without OOMing

Soft-mask inpainting with a strength dial

A UNET loader with the speed knobs built in

Getting a model's filename as text, not as a loaded model

Apply a LoRA to only the blocks you choose

Text-prompted masks, model auto-downloaded

Check your tracking data before you trust it

See your mask before you commit to it

Forcing a branch to run without saving anything

Your own list of canvas sizes

Stop memorizing SDXL resolutions

VAE-encode a video without loading every frame into memory first

Find out what's actually eating your VRAM

Skipping the base64/JSON tax on live previews

Preview a huge image batch without the lag

Drop the frames where detection found nothing

Reverse a sigma schedule for unsampling

One number, wired everywhere, changed in one place

Turn a list of numbers into a batch of solid masks

Build a custom noise schedule from a list of numbers

Turn off the parts of a LoRA you don't want

A noise latent you actually control

Slice frames out of a batch

Pull specific frames out of a batch by index

Read width, height, and batch size mid-graph

Slice a sub-range out of a latent batch

Pick specific latents out of a batch by index

Read a latent's real dimensions

The inline sanity check for mask batches

Save the frames from the denoise preview

Slice your tracking data down to the window you're sampling

Load a GGUF model with the extra knobs exposed

Animate GLIGEN boxes with a spline instead of typing coordinates by hand

Sample a gradient image into a numeric curve

The inpainting mask node you'll actually use

View HDR IC-LoRA output without it looking washed out

Dial in a Hunyuan LoRA block by block

Generate video that starts and ends on your frames

Shave VRAM off local Ideogram 4 runs (at your own risk)

Draw your prompt for Ideogram 4's JSON format

Blend more than two images without daisy-chaining nodes

See your mask on top of your image

Stitch video chunks without a hard seam

Scrubbing dead-black frames out of a batch

Stitch two clips together with a real transition

Merge several images into one batch, add inputs on demand

Hold each frame N times to stretch a batch

Number your frames so you can actually verify your batch logic

Stitch two images together, any direction

Turn a batch of images into one grid

Paste several images into one, in any direction

Crop an image down to whatever your mask covers

Zoom into a masked region for detail work

Tight crops around a mask, per frame

Pull a screen region straight into ComfyUI (local desktop only)

Four images into one comparison sheet

Nine images in, one contact sheet out

Slice a contact sheet back into frames

The little noise nudge for image conditioning

Rescale an image tensor when something downstream expects -1 to 1

Extend a canvas without touching your original pixels

Pad to a size for outpainting, no arithmetic

Add borders, letterbox, and set up outpainting

A do-nothing node that's actually useful for graph hygiene

Build the reference-plus-canvas image an in-context LoRA needs

The KJNodes resize node (and why you shouldn't use this one anymore)

The resize node that actually handles aspect ratio

Four real sharpening methods on the GPU, pick the right one

Batch images that aren't the same size

Make an image zoom and pan to the music

Crop, resize, pad, and rotate on one interactive canvas

Paste an inpainted crop back cleanly

Model upscaling that won't OOM on a batch

Add controlled noise back into a latent

Put processed frames back where they came from

Drop or replace frames at the indices you name

Drop a latent into a batch at a chosen index

One number, wired to as many places as you want

Smoothing a path of x/y points along a curve

Join many strings, or emit them as a list

Glue two strings together with a delimiter

Real token weighting for a Qwen3-encoded model

Turn a rough cut-and-drag edit into believable motion

A switch that actually skips the branch it doesn't pick

The community's fix for Hunyuan's missing image-to-video

Load, resize, and grab a mask in one node

Batch a whole directory into one workflow

Batch-load a whole folder of clips into frames

Pulling a LoRA out of a finetuned checkpoint

Shrink a LoRA's rank without retraining it

Tune LTX-2's audio/video attention (and trim VRAM)

Clean up LTX-2's generated audio

Dial LoRA strength per transformer block

Per-pathway LoRA strength for LTX-2's audio+video

Cut peak VRAM on LTX-2

Negative prompts for LTX-2, even at CFG 1

Get live previews back for LTX-2 video

Pin multiple keyframes into an LTX video

The AddGuide trick, wired up for a whole batch at once

Regenerate a time slice of an LTX-2 audio+video clip

Shave peak VRAM off LTX video

A one-dial detail boost for LTX video

Anchor LTX video to your input frames

Combine any number of masks into one batch

Turn a batch into a number schedule

Rebuild an image from separate R, G, B (and alpha)

Lie to ComfyUI's VRAM estimator

Watch VRAM move during sampling

A node that does nothing, on purpose

The fp16-accumulation toggle with a bad name

Full-resolution live previews during sampling

Bake a patched or merged diffusion model to disk

Experimental sparse attention for Kandinsky 5 video

Turning a soundtrack into numbers ComfyUI can use

Draw a shape that pulses to the beat

Shift, rotate, and batch-animate a mask

Make a mask shake and spin to the beat

Space out keyframes with gap frames

Force real flash attention, no silent fallback

A deprecated LoRA + torch.compile fix

The in-graph switch for a free speedup on Flux and Wan

Fixing Flux.2's brighten-and-blue habit

Get a sound when your queue actually finishes

Turn a coordinate list into an actual picture you can check

Click points and boxes onto an image for SAM and cropping

Play a frame batch in the graph without saving a file

One preview node for both images and masks

See the inpaint mask you're actually sampling with

Sampling frames without babysitting indices

Stretch or compress pixel values into a new range

Stretching or squashing a mask's intensity

Splice regenerated frames back into a batch

Match a mask to your image, without the gray fringe

Flip a batch of frames back to front

Turn a soft mask into a clean binary one

Fixing wacky video motion without extra training

Save images with captions and a folder you choose

Write a transparent PNG from an image plus a mask

Write text to disk like you'd save an image

Retime a prompt schedule to a new frame count

Apply CFG only over the steps that need it

Grab your screen straight into a workflow

Feed your live screen into ComfyUI as an image

Split one mask into per-object masks

Tell Shakker's Flux union which mode to run

Randomize the order of a batch, deterministically

Get your noise schedule out where you can actually see it

A real math expression node for your workflow

The SLG quality trick tuned for Wan

Pausing a ComfyUI workflow on purpose

Turn any value into text, no matter the type

Drive your generation from a live microphone

Draw a curve, get a schedule (or a motion path) out of it

Cut a detection list into two groups

Pull an image apart into R, G, B and alpha

Keyframe a camera orbit around an object

A proper VRAM allocator trace, not just a printout

A single-line piece of text you can wire around your graph

A text box you wire like any other node

Type a bunch of numbers, get a float list

The strength slider Flux Redux should have shipped with

Auto-expanding a short prompt into a detailed one

Scripting the camera orbit for Stable Video 3D

Time any slice of your ComfyUI workflow

Squeeze extra speed out of your ControlNet with torch.compile

Deprecated, and Cosmos moved on too

Deprecated, but here's what it did and what to use instead

The fine-grained torch.compile knob for video models

Deprecated — three names, one job

Torch.compile for Flux, block by block

Deprecated, from Kijai's own Hunyuan wrapper era

Deprecated — torch.compile for Qwen-Image now lives elsewhere

Deprecated, but the speedup it chased is real

Deprecated, use TorchCompileModelAdvanced

Torch.compile for just the VAE, with real backend and mode controls

Wipes and crossfades between stills

Crossfades and wipes between your frames

Kill the seam on a looping video

A VAE loader for pixel-space models

Blend two VAEs, if they're actually the same shape

Turn an OOM into a picture

See the noise schedule you're actually running

Force ComfyUI to free memory mid-workflow

Apply a LoRA to some of Wan's blocks, not all of them

Trade a little speed for a lot of VRAM on Wan

The KJNodes path to long Wan clips with SVI Pro

A near-free quality bump for Wan video

Cut Wan attention VRAM

Negative prompts for Wan when speed LoRAs force CFG 1

Deprecated, and what to use instead

Pull a live camera frame into your graph

Reshape a value curve for the next node

Stitch two value curves into one

Pull any node's setting out as text