FireLTX2.3_AIO

My ComfyUI workflow pack for LTX 2.3.

This is my own all-in-one workflow setup for generating and editing videos with LTX 2.3. It is built around the way I actually use LTX locally and is focused on good quality, practical controls, easier testing, and low-VRAM-friendly options.

This workflow was inspired by and modified from:

MrXin LTX 2.3 T2V - Eros - 12GB VRAM & 32GB RAM - Workflow

https://civarchive.com/models/2528728/mrxin-ltx-23-t2v-eros-12gb-vram-and-32gb-ram-workflow?modelVersionId=2943165

---

What this workflow is for

This workflow is meant for people who want a more practical LTX 2.3 AIO setup with cleaner controls and less guesswork.

Main goals

• Better quality

• Better motion stability

• Cleaner prompt testing

• Easier switching between generation modes

• Easier editing flow

• Better control over chunking / VRAM use

• Better support for prompt segmentation through LTX Director

• Practical settings for 10, 20, and 30 second generations

• Separate generation and editing stages to reduce VRAM problems

---

Supported workflow types

• T2V

• I2V

• FLF2V

• FML2V

• V2V

• Optional audio path

---

What you get

• Director-based prompting

• Global prompt

• Segment prompts

• Easier seconds-based timing

• Generation + editing support

• NAG support for negative prompting at CFG 1

• Chunking controls

• Tiled sampler controls

• Tiled VAE decode controls

• Latent upscaler support

• Preview VAE + final VAE options

• Cleaner layout for testing

• LoRA-ready

• I removed my own LoRA block so you can plug in whatever LoRAs you want

• Separate editing options for:

  • Anime

  • Realistic

  • 3D

• RTX Video Super Resolution

• Traditional image upscaling

• RIFE frame interpolation

---

Tested output and generation times

The workflow is currently tuned to generate at:

• 832 × 1536

• 24 FPS

My approximate generation times with the included optimized settings are:

| Video Length | Approx. Generation Time |

| 10 seconds | ~5.5–6 minutes |

| 20 seconds | ~11.5 minutes |

| 30 seconds | ~16.5 minutes |

These times are only a reference. Model choice, LoRAs, available VRAM/RAM, other GPU usage, ComfyUI version, and system configuration can change them.

10 second generation settings

• Chunking disabled

• First Pass Preview disabled

• LTX Tiled Sampler:

• n_tiles = 2

• tile_overlap = 4

• LTXV Spatio Temporal Tiled VAE Decode:

  • spatial_tiles = 4

  • spatial_overlap = 4

  • temporal_tile_length = 24

  • temporal_overlap = 4

20 second generation settings

• Chunking disabled

• First Pass Preview disabled

• LTX Tiled Sampler:

• n_tiles = 3

• tile_overlap = 4

• LTXV Spatio Temporal Tiled VAE Decode:

  • spatial_tiles = 4

  • spatial_overlap = 4

  • temporal_tile_length = 24

  • temporal_overlap = 4

30 second generation settings

• Chunking enabled

• chunks = 2

• dim_threshold = 4096

• First Pass Preview disabled

• LTX Tiled Sampler:

• n_tiles = 4

• tile_overlap = 4

• LTXV Spatio Temporal Tiled VAE Decode:

  • spatial_tiles = 4

  • spatial_overlap = 4

  • temporal_tile_length = 24

  • temporal_overlap = 4

30 seconds is the longest single generation I was able to run reliably with the included high-quality settings and LoRAs disabled.

If you add LoRAs and get an OOM, the chunking and tiled sampler settings are the first things I recommend adjusting.

---

Video Editor output

The editor is intended to be run after generation is finished.

Generate and save your video first, then upload the generated video into the editing section. This avoids keeping the large LTX generation models loaded while the editor is processing the video.

With the included interpolation/editing setup, my final edited output is:

• 720 × 1280

• 48 FPS

• Stereo audio when audio is present

The editor includes optional:

• Detail - Anime

• Detail - Realistic

• Detail - 3D

• RTX Video Super Resolution

• Upscale Model

• RIFE frame interpolation

Editing time warning

The full editing path is much slower than generation.

Using:

Detail - Realistic + RTX + Upscale

can take approximately:

30 minutes to 1 hour

depending on the length of the source video and the options enabled.

That is expected. If you do not need the additional restoration/upscaling, bypass those sections.

---

Required / Recommended startup arguments

These are the ComfyUI startup arguments I use with this workflow:

--enable-manager --reserve-vram 2 --preview-method none --async-offload --cache-lru 2

ComfyUI Desktop

1. Open ComfyUI Desktop.

2. Open Settings.

3. Find Startup Arguments.

Enter:
--enable-manager --reserve-vram 2 --preview-method none --async-offload --cache-lru 2

1. Restart ComfyUI.

What they do

• --enable-manager

  • Enables ComfyUI Manager.

• --reserve-vram 2

  • Reserves 2 GB of VRAM instead of allowing ComfyUI to consume everything available.

• --preview-method none

  • Disables ComfyUI's normal latent previews and saves some processing/VRAM overhead.

• --async-offload

  • Enables asynchronous model offloading.

• --cache-lru 2

  • Uses a small LRU cache.

These are the arguments I used while tuning the workflow. You may need to adjust them for your own hardware.

Required LTX Director patch

I made a small modification to LTX Director so the main Video Length (seconds) control can control Director while Director remains in seconds mode.

I specifically wanted to keep Director in seconds mode because placing timeline segments at 3 seconds, 7.5 seconds, etc. is much easier than manually calculating frame positions.

Without this patch, externally changing end_second / duration_seconds does not reliably update Director's internal duration_frames.

1. Find ltx_director.py

For ComfyUI Desktop it should be under:

ComfyUI/

└── custom_nodes/

└── WhatDreamsCost-ComfyUI/

└── ltx_director.py

A typical Windows ComfyUI Desktop path looks like:

C:\Users\<USERNAME>\AppData\Local\Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI\custom_nodes\WhatDreamsCost-ComfyUI\ltx_director.py

2. Back up the file

Before changing it, make a copy such as:

ltx_director.py.bak

3. Edit the execute() function

Find the execute() function in ltx_director.py.

Immediately inside the function, before the timeline parsing section, add:

# When using seconds mode, make the seconds inputs authoritative.
# This allows externally connected duration/end-second controls to
# correctly drive the internal frame values.
if display_mode == "seconds":
start_frame = int(round(float(start_second) * float(frame_rate)))
end_frame = int(round(float(end_second) * float(frame_rate)))
duration_frames = int(round(float(duration_seconds) * float(frame_rate)))

It should be placed before:

# Parse timeline data

4. Save and restart ComfyUI

Completely restart ComfyUI after changing the file.

The workflow can then use:

Video Length (seconds)

↓

Int To Float

├──→ LTX Director.end_second

└──→ LTX Director.duration_seconds

while Director remains in:
Display Mode = seconds

Important

Updating or reinstalling WhatDreamsCost-ComfyUI / LTX Director may overwrite this modification.

If the main Video Length control suddenly stops changing the actual generated duration after updating Director, check ltx_director.py and reapply the patch.

Main notes

• This workflow is built around the models listed in the resource description.

• File paths may need to be adjusted depending on your local setup.

• If something is not loading, check:

  • model folder location

  • file names

  • node pack installation

  • ComfyUI version

• If audio output is bad or static-heavy, disable audio generation and render silent video instead.

• Some settings are meant for quality, some for speed, and some for longer clips. Do not change everything at once.

• Generation and editing are intentionally separated to reduce VRAM usage.

• LoRAs can increase memory usage and may require more aggressive chunking or tiling.

Known issues / advice

• Do not use ComfyUI Nodes 2.0 beta with this workflow.

• Install ffmpeg.

• Keep ComfyUI, PyTorch, and your custom nodes updated.

• Make sure your model paths match your own system.

• Some custom node packs may still require manual installation depending on your setup.

• If you get OOM:

  • enable or increase chunking

  • increase the number of LTX Tiled Sampler tiles

  • reduce duration

  • reduce resolution

  • disable extra passes / extras one at a time

• Audio generation can be inconsistent depending on prompt and model choice.

• The full video editing path can take considerably longer than generation.

• Updating LTX Director may overwrite the required seconds-mode patch.

Credits

Big credit to MrXin for the workflow that helped inspire the starting point for this setup.

This version is not a reupload of that workflow. It is my own modified setup built around my own testing, layout changes, generation flow changes, Director support, NAG support, quality tuning, VRAM optimization, speed optimization, editing changes, and general cleanup.

Original workflow:

https://civarchive.com/models/2528728/mrxin-ltx-23-t2v-eros-12gb-vram-and-32gb-ram-workflow?modelVersionId=2943165

Final note

You are responsible for what you generate with it.

If you find the workflow useful, a like is appreciated.