CRT-Nodes

Bloom, plus a full color-grading suite you probably don't need yet

Turn 'something is plugged in' into a boolean

A glow filter with actual controls

A 'tube' compressor for generated video audio

Make an audio clip match your video's frame count

Grab a random audio file from a folder, deterministically

Make your video exactly as long as your soundtrack

Actually hear the audio in your ComfyUI graph

Sanitize autogenerated prompts before they wreck your render

A U-shaped fade across an image batch

A NOT gate for your graph

Turn a string into an INT and a BOOLEAN

Mid-tone contrast that doesn't wreck your edges

Keep one hue, kill the rest

Saturation that won't blow out your highlights

Turn a photo into stylized outlines

Stamp the prompt and settings onto the image, like a camera info plate

Point it at a folder of audio, get a padded batch — with the filenames riding along

Batch transcription that actually pairs each transcript with its file

The text encoder ChronoEdit forgot to tell you about (and this node downloads for you)

The vision encoder for ChronoEdit's image-to-video trick (auto-downloaded)

This LoRA turns 50 steps of editing into a handful

This node downloads the fp8 weights for you

Bigger edits without smearing them

ChronoEdit uses Wan's VAE — this node fetches it so you don't have to

The text encoder for Baidu's infographic model

Auto-download Baidu's ERNIE-Image diffusion model

The ERNIE Turbo model node that downloads itself on first run

The one VAE you need for ERNIE-Image

Auto-download Klein's text encoder

The HDRI 360 panorama LoRA for Flux Klein, downloaded for you

Flux 2 Klein in ~10GB, downloaded on first queue

Auto-download Klein's VAE

The Krea 2 text encoder, without the 4GB guesswork

The undistilled base, one click away

The 8-step Krea 2 checkpoint, auto-downloaded

One node, zero setup, the right VAE

The audio half of LTX2.5's split VAE, downloaded for you

LTX2.5's 12B text encoder, w4a8 quantized, auto-downloaded

A model patch for variable-length generation

The union-control LoRA that gives LTX2.5 Canny-style conditioning

Bigger frames, learned detail

The LTX2.5 model node with four VRAM tiers that downloads itself

LTX2.5's outpainting IC LoRA

LTX2.5's latent spatial upscaler, 2x, downloaded on first run

The half that doubles the frame count

Detail from the model itself, not a post-pass

LTX2.5's video VAE, self-downloading, zero configuration

The audio half of H3's two-VAE design, self-downloading

The 32B text encoder H3 needs, INT8 or NVFP4, downloaded for you

MiniMax H3's FL2VA Turbo LoRA — 4 steps, and the node fetches it for you

Two families, four quants, zero hunting

REF2VA Turbo LoRA for MiniMax H3 — 4-step reference-to-video

H3's video VAE, no inputs, one download, done

Clean cutouts before the 3D even starts

The eyes Pixal3D sees through

The Pixal3D diffusion model, minus the hunt for the right file

The camera that figures out your FOV

The geometry half of Pixal3D's two-stage brain

The half of Pixal3D that paints your mesh

The segmentation workhorse behind SAM3D Body

Full 3D body mesh from a single photo

Same MoGe camera, different shoot

The person detector that sets up the body scan

Z-Image Turbo's text encoder, with the download handled

Z-Image Turbo, zero manual downloads

The Z-Image VAE that downloads itself

Green screen keying that lives in your node graph

Free your VRAM the moment your prompt is encoded

DA3 that downloads its own weights

One conditioning output that encodes a whole batch of prompts

Thin out a frame batch evenly — the video decimation node

Run a folder of text prompts as a batch, and get the conditioning out

Run a whole prompt folder through Krea 2 with phrase weights — one node, CFG 1.0

The two-frame node behind every Wan extension chain

Grab a deterministic window of images from a folder, seed by seed

An integer, but as a node — the boring utility that unblocks your graph

Track and crop a face through a whole video with SAM 3.1

Text-prompted cropping via CLIPSeg

Composite your enhanced subject back onto the original canvas

The two-string glue node you'll outgrow (but want anyway)

Batch sampling that stops re-encoding your prompts every image

The batch sampler with a two-stage detail daemon bolted on

Grab the newest video in a folder without typing its name

One node, the entire LTX 2.3 pipeline — audio, depth, upscale and all

The config half of the LTX 2.3 one-stop sampler

Bundle every LTX 2.3 model into one pipe for the unified sampler

The node that runs all of MiniMax H3 — T2V, I2V, R2V and audio, in one place

MiniMax H3 US Config — one prompt box to drive the whole generation

Four MiniMax H3 models, one socket — and it only loads what the run needs

MiniMax H3 won't take just any frame count — this node snaps yours to the grid

Crop by percentage, not by pixel math

One node that can color-grade, sharpen, glow, grain, and vignette

Snap any image to a training-ready resolution

Chop lines off your prompt lists without editing by hand

Turn N prompt fragments into one batch-ready string

How many prompts are actually in that text block?

Turn one block of text into the shots of your video

A text box that shows you the string and lets it pass through

Watch Your Model Think Out Loud — the CRT Node With No Inputs That's Somehow Useful

One node to grab the last frame of a video latent without decoding the whole thing

The whole two-model Wan pipeline in one node

Fix TensorRT depth output so ComfyUI can actually use it

Enable Latent (CRT)

Score every image 0-100 in one pass — the ERNIE aesthetic filter that curates for you

The node that turns a JSON Q/A pair into a prompt an LLM can finish

Ultralytics Enhancer (CRT)

SEGS Enhancer Multi (CRT)

The note node that doesn't look like a gray box

The stopwatch that lives on your canvas

Give your renders the filmic texture they're missing

Dial your Flux LoRA up by block, not as a single strength

Size your latent to a pixel budget, not a guess

The megapixel sizing node with the 64-grid for video models

Load every image in a folder one at a time, without touching the graph

Keep every image in a batch between two megapixel bounds

Check your tile seams at a glance with Image Tile Checker (CRT)

A KSampler with a built-in detail fix

Barrel, pincushion, anamorphic — fake the lens, keep the film look

Chromatic aberration, vignette, and film grain in one node

How Photoshop hands ComfyUI a layer without saving a file

Load Image Resize (CRT)

Load Last Image (CRT)

Load Last Latent (CRT)

Re-open a saved latent + conditioning pair by seed — and grab the matching audio for free

A LoRA loader that merges, stacks, and hands you the merged weights

Turn a merged LoRA into a file you can actually ship

Auto-blur or pixelate faces before your AI video goes public

Did that mask actually find anything? Ask it, get a number

Mask Pass or Placeholder (CRT)

Mask Temporal Enhancer (CRT)

Slotting the LLM's answer back into the JSON

Your mono voice track needs to be stereo. This makes it happen

An eight-band parametric EQ, right inside your node graph

Image Upscale Sampler (CRT)

Reference Latent Batch (CRT)

Pick a megapixel target and get the right latent for your model

Stop guessing what resolution your model actually wants

Sweep every sampler and scheduler combo by bumping one number

Turn sampler and scheduler names into wireable strings

A save node that actually lets you pick the folder

For when your workflow's output needs to leave as data, not a file

Save images exactly where you want them, without ever overwriting

Save JPEG Websocket (CRT)

Freeze a latent and its conditioning into one file, so you can re-run the exact same scene later

Park a latent on disk and pick it up later

Save Text With Path (CRT)

Save your video where you actually want it, not in the output swamp

Scale to a resolution target, not a pixel size

Seamless Loop Blender (CRT)

A float slider that looks like a dial

The T node is just a boolean switch — and that's the point

An edge-aware smoothing filter (not the sampler's denoise)

A blank canvas node, and why you actually need one

Strength to Steps (CRT)

That saturated old-film color-grade look, in one node

Append to .txt without ever overwriting

Text Box line spot (CRT)

Text Loader Crawl (CRT)

Run a whole folder of prompts without retyping a thing

Walk a list one item per run with ComfyUI's own seed widget

Talk to whatever model is loaded in Unsloth Studio, straight from your graph

Video Duration Calculator (CRT)

Load a video from a folder by seed — frames and audio in one shot

Same seed, same prompt, every LoRA laid out side by side

Separate LoRA stacks for the high and low noise passes