LTXV 2.3 First Frame to Last Frame ComfyUI Workflow + One Click Installer
The LTX 2.3 First Frame and Last Frame workflow introduces a more story-driven image-to-video generation process by allowing full control of both the starting and ending frames. This enhancement transforms standard AI-generated clips into visually consistent sequences with smoother scene progression and more cinematic transitions.
I came across the original workflow on the Whatdreamcosts GitHub repository and have made a few performance-focused tweaks to improve compatibility with GGUF UNet and CLIP models. These adjustments make the workflow significantly faster and better suited for low VRAM devices.
In testing on a 24GB RAM GPU, I achieved a 1280×640 10-second video in about 3–4 minutes using all three processing phases. You can explore the original workflow and a full video tutorial by the creator here:
Whatdreamcosts GitHub Repo & Tutorial:
https://www.youtube.com/watch?v=aXDIr8eNovI
One Click Installer (Low VRAM GGUF Version)
To make setup effortless, I’ve built a One Click Installer for the lower VRAM GGUF version of the LTX 2.3 workflow. It includes all the essential model files, preloaded in the correct ComfyUI directory structure so you can start generating immediately.
Preloaded Models Within the Installer (Low VRAM):
ltx-2.3_text_projection_bf16.safetensors – Hugging Face Link
gemma-3-12b-it-UD-Q5_K_XL.gguf – Hugging Face Link
LTX23_audio_vae_bf16.safetensors – Hugging Face Link
LTX23_video_vae_bf16.safetensors – Hugging Face Link
ltx-2.3-22b-dev-Q3_K_S.gguf – Hugging Face Link
ltx-2.3-spatial-upscaler-x2-1.0.safetensors (Upscale Model) – Hugging Face Link
ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors (LoRA) – Hugging Face Link
The standard LTX 2.3 22B checkpoint models (BF16, FP8) aren’t bundled within the installer, but can be easily downloaded from Kijai’s official Hugging Face repository:
LTX-2 19B Diffusion Models
Speed & Performance
You can generate 10-second 1280×640 video in about 3–4 minutes on 24 GB VRAM.
The workflow intelligently adapts to your hardware—running efficiently on lower VRAM GPUs while still achieving smooth motion and crisp visual detail. The process begins at lower resolution and automatically upscales 2× in the final two sampling phases to deliver stunning, high-resolution renders.
System Requirements
NVIDIA RTX 30XX / 40XX / 50XX GPU (FP16 supported)
CUDA-compatible GPU (minimum 20GB VRAM, 24GB+ recommended)
Windows OS
Minimum 50GB free storage
What’s Included
Portable ComfyUI Windows Installer (pre-configured for LTX 2.3 image-to-video)
Automated model downloads and ComfyUI nodes setup
Optimized long-sequence video workflows for flexible use
Designed for both beginners and advanced AI video creators
Usage Notes
Fully compatible with LTX 2.3 22B GGUF (optimized for Low VRAM setups).
Use rich, detailed text prompts for better motion coherence and narrative control.
For the First Frame / Last Frame workflow, start with high-resolution reference images to ensure clarity and consistent detail throughout your video.
Add both images in the Multi Image Loader, arranging image 1 as the start and image 2 as the final frame.
Adjust resolution based on your GPU’s capacity—higher resolutions demand more VRAM.
To customize or expand the workflow, open the subgraph’s prompt node and use the top-right arrow for advanced settings.