Extensions/comfyui-extend
ComfyUI Extension

comfyui-extend

Small ComfyUI custom nodes for building a Flow-style video continuation workflow with LTX 2.3.

By balarooty·Created 4 months ago·Updated 3 months ago· 0
balarooty/comfyui-extend
Nodes—
On cloudLocal install
Stars0
Updated3 months ago
Readme

ComfyUI-LTXFlow

Small ComfyUI custom nodes for building a Flow-style video continuation workflow with LTX 2.3.

This package does not replace the official LTX or Qwen nodes. It adds the missing workflow glue:

  • extract a specific frame from a generated clip
  • extract a final tail segment for better continuation
  • stitch multiple clips into one frame batch

Install

Place this folder in:

ComfyUI/custom_nodes/ComfyUI-LTXFlow

Then restart ComfyUI.

Download Models

From this repo:

COMFYUI_DIR=/workspace/ComfyUI bash scripts/download_models.sh

Change COMFYUI_DIR to your actual ComfyUI path. The script downloads every model needed by:

workflows/05_qwen_edit_to_ltx_first_last.json

It supports Hugging Face auth if needed:

HF_TOKEN=hf_xxxxx COMFYUI_DIR=/workspace/ComfyUI bash scripts/download_models.sh

This script only downloads model files. It does not install ComfyUI custom node packs.

Install Workflow Custom Nodes

The Qwen Edit to LTX workflow also depends on custom node packs outside this repo. If ComfyUI shows an "Installation Required" dialog for workflows/05_qwen_edit_to_ltx_first_last.json, install these packs with ComfyUI Manager:

  • ComfyUI-KJNodes - provides LazySwitchKJ, VAELoaderKJ, and LTX2SamplingPreviewOverride
  • WhatDreamsCost-ComfyUI - provides LTXSequencer
  • ComfyUI Easy Use - provides easy mathInt

Manual install:

COMFYUI_DIR=/workspace/ComfyUI bash scripts/install_custom_nodes.sh

Change COMFYUI_DIR to your actual ComfyUI path. The helper clones or updates:

Restart ComfyUI after installing custom nodes, then refresh the browser page.

Frontend

Custom frontend code lives here:

js/ltxflow_ui.js

The package exposes it through:

WEB_DIRECTORY = "./js"

This frontend adds node colors, short in-node notes, and default widget adjustments for LTXFlow nodes. The main node interfaces still come from each Python node's INPUT_TYPES, which is how ComfyUI builds standard node widgets.

Nodes

Proper LTX 2.3 first/last workflow, based on the WhatDreamsCost reference workflow:

workflows/04_ltxflow_first_last_wdc_ltx23.json

Use this one for the first real test. It requires the same dependencies as the reference workflow:

  • WhatDreamsCost-ComfyUI
  • official LTX/ComfyUI LTX nodes used by that workflow
  • VideoHelperSuite SaveVideo/video combine dependencies
  • this ComfyUI-LTXFlow package

Qwen Image Edit to LTX first/last workflow:

workflows/05_qwen_edit_to_ltx_first_last.json

This is the main Qwen + LTX workflow. It wires:

source image + reference image
  -> Image Edit (Qwen-Image 2511)
  -> LTX Flow - Qwen Edit Bridge
  -> LTX Flow - First/Last Guide
  -> LTX sampler/video output

The bridge sends the original source image as the LTX first frame and the Qwen-edited image as the LTX last frame.

Required custom node packs:

  • ComfyUI-KJNodes for LazySwitchKJ, VAELoaderKJ, and LTX2SamplingPreviewOverride
  • WhatDreamsCost-ComfyUI for LTXSequencer
  • ComfyUI Easy Use for easy mathInt

Required Qwen models from the official ComfyUI template:

  • qwen_2.5_vl_7b_fp8_scaled.safetensors in ComfyUI/models/text_encoders/
  • Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors in ComfyUI/models/loras/
  • qwen_image_edit_2511_bf16.safetensors in ComfyUI/models/diffusion_models/
  • qwen_image_vae.safetensors in ComfyUI/models/vae/

Required LTX 2.3 models:

  • ltx-2.3-22b-dev-fp8.safetensors in ComfyUI/models/checkpoints/
  • gemma_3_12B_it_fp8_scaled.safetensors in ComfyUI/models/text_encoders/
  • ltx-2.3-22b-distilled-lora-384.safetensors in ComfyUI/models/loras/
  • ltx-2.3-spatial-upscaler-x2-1.1.safetensors in ComfyUI/models/latent_upscale_models/

You do not need ltx-av-step-1751000_vocoder_24K.safetensors for this workflow. That was an older/reference workflow name, not the current LTX 2.3 filename we use here.

The workflow uses the official Qwen Image Edit 2511 subgraph node from ComfyUI's workflow templates.

The guide section is:

MultiImageLoader
  -> LTX Flow - Extract Frame (index 0)
  -> LTX Flow - First/Last Guide first_frame

MultiImageLoader
  -> LTX Flow - Extract Frame (index -1)
  -> LTX Flow - First/Last Guide last_frame

LTXVConditioning + video VAE + EmptyLTXVLatentVideo
  -> LTX Flow - First/Last Guide
  -> sampler path

Combined workflow:

workflows/00_ltxflow_utilities_combined.json

This puts Extract Frame, Extract Tail, and Scene Builder on one canvas. Use it as the main starting workflow.

LTX Flow - Extract Frame

Input: video_frames

Output:

  • frame
  • resolved_index

Use index = -1 for the last frame, 0 for the first frame, or any positive/negative frame index.

Workflow:

workflows/01_extract_frame.json

LTX Flow - Extract Tail

Input: video_frames

Output:

  • tail_frames
  • actual_frame_count

Use this before an LTX continuation workflow. For best continuation behavior, start with tail_length = 17. Good values are 9, 17, and 25.

Workflow:

workflows/02_extract_tail.json

LTX Flow - First/Last Guide

Input:

  • positive
  • negative
  • vae
  • latent
  • first_frame
  • last_frame

Output:

  • guided positive
  • guided negative
  • guided latent

This is the important node for Frames-to-Video style generation. It uses ComfyUI's LTXVAddGuide internals to inject the first frame at frame 0 and the last frame at frame -1.

LTX Flow - Qwen Edit Bridge

Input:

  • source_image
  • qwen_edited_image

Output:

  • first_frame
  • last_frame
  • keyframe_batch

This node resizes the source and Qwen-edited image to matching dimensions, enforces a configurable multiple such as 32, and prepares them for LTX Flow - First/Last Guide.

LTX Flow - Tail Guide

Input:

  • previous clip tail frames
  • LTX conditioning
  • video VAE
  • target latent

Output:

  • guided positive
  • guided negative
  • guided latent
  • overlap_frames

This is for the next step: Flow-style extension. Use Extract Tail from the previous clip, feed it into Tail Guide, generate the next clip, then trim the overlapped frames before stitching.

LTX Flow - Trim Frames

Removes frames from the start/end of a generated batch. This is needed after overlap-based extension, because the first generated frames repeat the previous clip's tail.

LTX Flow - Scene Builder

Inputs:

  • clip_1
  • optional clip_2 through clip_6
  • deduplicate_joins
  • fps

Output:

  • merged_frames
  • total_frames
  • duration_seconds

Send merged_frames to VHS_VideoCombine or another video saver node.

Workflow:

workflows/03_scene_builder.json

Recommended Full Pipeline

Qwen Image / Qwen Image Edit
  -> keyframe image
  -> official LTX 2.3 image-to-video workflow
  -> LTX Flow - Extract Tail
  -> official LTX 2.3 continuation / multi-condition workflow
  -> LTX Flow - Scene Builder
  -> VHS Video Combine

Qwen is best used for first frames, last frames, character references, and visual consistency. LTX is best used for motion and clip continuation.