VRGameDevGirl Video Enhancement Nodes

Find the beats in your mix so your video cuts on them

Turn a beat map into SRT scene timings that actually cut on the beat

Make Every Shot Look Like It Was Graded Together

Film Grain Without Murdering Your VRAM

The Aggressive Sharpen for Detail Recovery — Use Sparingly

Directional Sharpening for When You Need Edges, Not Halos

The Sharpen You Actually Want for Video

Load image number N from a folder — the loop-friendly way

Re-load the exact reference frame by matching the filename number, not the sort order

Attach real durations to your lyric segments so prompts know their time budget

Housekeeping for your LLM batch runs, so the next one starts clean

00'), not raw seconds

Crop audio by plain float seconds — the pipeline-friendly trim

Stagger audio per chunk with an offset — the sync fix for chunked generation

It's a sticky note. One label in, nothing out — and that's the whole point

A Little String Node That Decides Where Your Video Gets Saved

Build a per-chunk output filename — and decide whether old takes get backed up

How many 62-second runs does your song need? This node tells you — and says it out loud

The queue-aware version of the set math — it hands other nodes everything they need

ChatGPT image generation inside ComfyUI — no API key, just a real browser

The normalization step this whole pack assumes — 48kHz, stereo, frame-aligned

One click to clear ComfyUI's memory — the tiny node that saves your video runs

Stitch up to 16 scene clips into one video, each trimmed to its song duration

Stitch 16 Scene Videos Into One Timed Sequence

The combine node that knows which set it's on — and labels each scene for you

Wait until enough final videos exist, then load them all as one clip

Concatenate all the set videos and glue the clean audio back on — the last node

The final-assembly node that also handles redo runs without clobbering your good take

A silent audio track, on demand, for the gaps in your LTX-2 timeline

One text list, five ways to pick from it

A node that just shows you the number — that's the whole point

Grab one duration out of the list

Twenty cycling text pickers, one friendly face

Turn one frame number into a list, and let the frames ride along

Sew the repaired face back into the video

The smarter Face Fix composite — aligns eyes, nose, and mouth before pasting

Paste the repaired face back — the no-nonsense composite node

Stash a review copy of your face-fix crop — so you can actually watch it

The Bridge That Hands Your Face Fix to LTX Video

Feeding your enhanced anchors to LTX the VHS way

The gate that won't let LTX run before it's ready

The node that finds a face, tracks it, and preps it for repair

The node that turns a whole video into a fixable face — shot by shot

The stripped-down Face Fix prep — shot-aware tracking, no Z-Image anchors

Turn the repaired face frames into an mp4 you can actually watch

Saving your enhanced anchor faces so LTX can steer by them

Google Flow inside ComfyUI — send images in, get a result back, no API key

The one-time setup that makes the VRGDG browser nodes work

A local LLM inside ComfyUI that writes your scene prompts

Rendering 400 scenes without babysitting the queue

A Vision-Language Model in Your Graph, With 24 Image Slots

The node that just knows where your audio file lives

Extract a clean filename prefix from a folder path without thinking

The next number in the sequence, computed for you

Your run number, straight out of the workflow's JSON state

Split a Song Into Vocals, Drums, Bass, and Everything Else

A pass-through node that nags you (and you can turn the nagging off)

Batch-load images by path without wrestling dynamic sockets

Before/after, right on the canvas — no file saving required

When the index hits 0, hand the edit an empty canvas

An image switch with a lookup table instead of a rule

Paste an enhanced face crop back in without the tell-tale hard edge

Pick one, two, or four images with a comma string

Up to 50 images on one switch, with a Refresh Inputs button

The multi-switch that gives index 0 a blank 1024×576 canvas

One Giant Prompt, Fifty Scene-Sized Prompts, Zero Copy-Paste

Same Prompt Splitter, But Now It Knows When to Re-Run

The Two-Second Type Fix That Unsticks Stubborn Nodes

The 5-second node every text pipeline ends up needing

Turning a JSON blob into something an LLM can actually read

Ostris' AI Toolkit, installed in an isolated venv — for the experimental Krea 2 edit path

The Krea 2 training cockpit — chunk training, sampling, and compare grids in one UI

Train a chunk of a Krea 2 LoRA and get a native safetensors path back

Stand up a Krea 2-ready Musubi-Tuner without the manual dance

Grab the newest subtitle file without touching a file picker

Llama-cpp-python Broke Again? This Node Diagnoses It

The multi-provider LLM brain behind the whole music-video pipeline

Keep every LLM prompt it generated — across a long, chunked run

Automatically write a scene prompt per beat of your story, in batches

Always point at the audio you most recently dropped in

Split audio by hand-tuned scene lengths, not a fixed grid

The workhorse chunker that turns one song into a queue of scenes

The plainest HUMO scene splitter in the family

Split scenes and grab the lyrics in one pass

The per-set transcription splitter that takes context prompts

Transcribe, set up scenes, auto-queue

Split scenes where the subtitle file says the lyrics change

Slice an Audio Track Into Scene-Sized Chunks, Right in the Graph

The Audio Splitter That Auto-Queues Its Own Scenes

Pick a file, get the audio, the path, and a clean name back

Pull the Latest Batch Results as Text, Without Touching the Filesystem

Load a pre-written prompt file, sorted by what it's for

Read back anything the workflow generated, from lyrics to style

Point It at a Folder, Get Frames You Can Actually Run

A local LLM node that needs no API key (even though it has a key box)

The LoRA dataset builder that lives inside the pack — no separate app needed

Apply a LoRA straight from a file path — no copying into the loras folder

Feed up to four subjects and a background to LTX 2.5's Multi-Reference Guide

Train an LTX-2.3 voice LoRA from inside ComfyUI — no command line

Train an LTX 2.3 Audio-Video LoRA From a Single Short Clip

One CFG value per denoise step, instead of one for the whole video

Pin both ends of the clip and let the middle figure itself out

A start frame, an end frame, and a smooth handoff between them

The Reference Sheet Builder That Keeps Your Character From Drifting

Train an LTX-2 LoRA From Inside ComfyUI, One Chunk at a Time

Build the multi-subject reference sheet in-node, not in a paint app

Turn your LoRA training previews into one comparison grid video

A CFG Guider that actually follows a schedule

CFG, STG, and variance rescale in one guider, scheduled per step

Let the image guide go so the video can move

Film-grade color grading on your video frames with a dropdown

Glue SRT timings onto your lyric segments so scenes stay in sync

When Your LLM Emits Broken JSON, This Is the Firefighter

Turning 'Na Na Na Na' Into Something a Video Model Can Use

Giving Every Lyric Line a Mood Before It Hits the Model

Three Hex Colors Into a Cinema Look

Give It a Song, Get Back Per-Scene Lyrics

Turn an SRT subtitle file into scene-sized lyric segments

Turn audio into word-timed lyrics your video builder can actually use

The lyric extractor variant the v9 builder workflow actually shipped with

No API key, no token — it just drives Chrome

Make MiniMax H3 keep your song instead of inventing one

The loader for MiniMax H3's learned latent upscaler

Upscale the video before it ever becomes pixels

Point it at your reference files and let it fill the H3 sockets

Splice the big video back into the latent without losing the audio

A LoRA loader that knows about pruned H3 checkpoints

Find the face in a wide shot, crop it cleanly, and remember where it was

Twenty Prompt Roulettes in One Node, for When Scenes Shouldn't All Match

Stack a pile of reference images and condition on all of them

Type the file paths instead of dragging fifty sockets

Twenty Text Boxes, One Prompt, No Suffering

Mixing Two Music3 Acoustic Plans to Find the Song in Between

The Volume Knob for Music3's Hidden Acoustic Plan

Because Music3 Will Happily Sing Your Metadata

Fill In the Blanks, or Let the Preset Do It

Four Deterministic Seeds So You Stop Comparing Apples and Oranges

Pick a Behavior, Not a Bunch of Numbers

The assembly line for a music video, audio in and scene plan out

Turn lyrics and a style brief into a cinematic scene prompt

Turn Lyrics and a Mood Board Into Scene Prompts

The Brain of the Music-Video Workflow (Yes, It Needs a Gemini API Key)

Get Musubi-Tuner and Its Models Without Leaving ComfyUI

Mute Node Groups Mid-Run

Stage Two of the Self-Running Workflow

Where the Batch Run Hands Back to You

Google's Flagship Image Model, Wired Straight Into Your Graph

The sticky note in your VRGDG workflow (it does nothing, on purpose)

Stack Twenty LoRAs, Safely, Without Breaking Someone Else's Workflow

When Half-Strength-First Isn't Right for Every LoRA

Freezing the last frame so your video doesn't jitter

This Node Does Nothing — and That's the Point

A control panel for workflow part 3

Stepping scene-by-scene across runs

The Prompt Creator's control surface, minus the plumbing

The current Prompt Creator, and why there are two of them

How this pack keeps a character consistent

When your prompt map comes back broken, this repairs it

Pick scene prompt N from a JSON list, without touching the graph

Splitting one big prompt into N scene prompts, the original way

When your prompt map is tiny

The four-scene splitter, for when a section is exactly four beats

Pick one scene from a prompt map by index

The scene-picker splitter that defaults to an array

Preview one scene at a time, by hand

The one output that works anywhere

The JSON-aware splitter with a summary port on the side

Sixteen outputs, no index, text straight in

Building a structured scene brief, section by section

A safe Python sandbox inside ComfyUI, with the sharp edges off

A queue trigger that marches the graph through audio-scene sets

A local Qwen LLM that writes your scene prompts

The brain of the VRGameDevGirl music-video machine

A Qwen brain inside ComfyUI, no API key, no Ollama server

The node that feeds your whole music video one scene at a time

A breadcrumb trail for multi-scene video renders

Put processed audio back on disk, where it belongs

Save your audio, and leave a note saying where you put it

Where the storyboard keeps its notes

Save text where you actually want it

The text-saver that grows a JSON list as you go

A mute/bypass dashboard for up to twelve workflow groups

Flip a batch of nodes off (or bypass them) from inside the graph

A debug display that takes literally anything

A terminal node that parks your image in the output folder

See any string in your graph, no wires left behind

Split a block of text in half, without splitting a word

Read the emotion curve out of your audio track

Train an LTX character LoRA without leaving the graph

Split one scene prompt into its image and motion halves

The one-click batch pass that makes clips look finished

One button, and the storyboard builder opens itself

Plan every scene prompt before you spend a single GPU hour

The node that queues your whole storyboard, one scene at a time

When your LLM hands you broken JSON, hand it back

Turn an LLM's text into something your graph can parse

Assemble the master music-video prompt from its parts

Run Gemma as a GGUF inside ComfyUI, no llama.cpp shell needed

Your concepts file, rewritten into shoot-ready video prompts by a local Gemma

The most boring node in this pack, and you'll use it constantly

One block of ideas in, ten prompt ingredients out

The tiny math node that keeps scenes in sync

Turn any song into timestamped lyrics your video builder can use

Turn a Song Into Lyric Text, No External App Needed

A counter that walks your scenes so every run generates something new

Cut the finished video exactly to the music, without opening an editor

Render one scene at a time, not the whole video at once

Slice frames to match your subtitles, so the lipsync actually lines up

The VRAM janitor for the pack's Gemma/GGUF prompt brains

Batch-edit every prompt in your music video without touching a JSON file

The same prompt updater, tuned for Z-Image's still-image pass

A scratch canvas for Video Builder ideas, parked inside your workflow

The before/after wipe that settles whether your enhance pass actually helped

Pull a single scene out of the editor session, ready to re-render

Edit scenes, flag remakes, generate prompts from a frame

The merge point that proves both enhance branches finished before LTX runs

Feed the Z-Image enhance pass one anchor at a time through VHS Meta Batch

Slice your video, choose anchors, plan the LTX pass

Put the enhanced video back to the exact size it started as

Save the Z-Image-enhanced anchors in the order LTX will need them

Put a wall of videos side by side and judge them at a glance

Chop a long video into LTX-sized chunks so the model can chew them

Local Voice Cloning Inside ComfyUI, With a Modes Menu

Train a Z-Image character or style LoRA in resumable chunks, ComfyUI-native

Train a Z-Image character LoRA without ever leaving ComfyUI

Run the bundled Z-Image text-to-image workflow from a canvas, not a file browser