Workflow Goal

Provide a clean, educational reference implementation for LTX 2.3 Text-to-Video generation.

This workflow intentionally prioritizes readability, reproducibility, and educational value over maximum performance or feature completeness. It serves as a stable reference implementation that users can understand before adding their own optimizations.

Treat this workflow as a starting point, then customize the prompts and generation settings to match your preferred artistic direction.

Version

v1.0

Purpose

• ✅ Educational Reference Workflow

• ✅ Clean Baseline Implementation

• ✅ Dynamic VRAM Friendly

• ✅ Tiled VAE Decode

Features

• ✅ Wireless Modular Architecture

• ✅ Clean Reference Implementation

• ✅ Prompt Engineering Ready

• ✅ Two-Phase Sampling Pipeline

• ✅ Low VRAM Compatible

• ✅ Modular and Easily Extensible

Workflow Pipeline

Engine Configuration
        ↓
Video Clip Configuration
        ↓
Image Preprocessing
        ↓
Prompt Engineering
        ↓
Image Conditioning
        ↓
Phase 1 Sampling
(Base Latent Generation)
        ↓
2× Spatial Upscale
        ↓
Phase 2 Sampling
(Detail Refinement)
        ↓
VAE Decode
        ↓
Final Video Output

Verified Settings

The following settings were used to validate this workflow and are recommended as the baseline configuration.

Sampling

• CFG: 1.0

• Steps: 11 (8 + 3)

• Sampler: Euler Ancestral

Video

• Frames: 145 (≈6 seconds @ 24 FPS)

• Output FPS: 24

• Tested Resolution: 1024 × 576 (16:9)

Required LoRA Configuration

No LoRAs were used during validation.

This workflow has been validated using the LTX 2.3 Distilled GGUF model. The official Distilled LoRA is not required when using the distilled model and has therefore been omitted from this reference workflow.

Hardware & Resolution Notes

Validated On

Hardware

• GPU: NVIDIA RTX 2080 (8 GB VRAM)

• System RAM: 64 GB

Software

• ComfyUI: 0.27.x

• Python: 3.13.7

• PyTorch: 2.12.0 + CUDA 13.0

Validation Settings

• Resolution: 1024 × 576

• Length: 6 Seconds

• Generation Time: ~12–14 Minutes

Although compatible with 8 GB VRAM, LTX 2.3 Text-to-Video remains computationally intensive and memory hungry.

System RAM plays an equally important role in handling intermediate tensors and memory paging during video generation.

Expected Behavior

8 GB VRAM + 16 GB RAM

• Possible out-of-memory errors

• Slower generation

8 GB VRAM + 32 GB RAM

• Better stability

• Performance depends on available system memory

8 GB VRAM + 64 GB RAM (or more)

• Recommended configuration

If your system has a more capable GPU, consider increasing the output resolution to 720p or higher.

Design Philosophy

This workflow intentionally avoids unnecessary complexity.

The objective is to provide a stable, reproducible reference implementation that users can understand, learn from, and extend for their own creative projects.

Advanced optimizations such as SageAttention, Flash Attention, or other hardware-specific accelerations have been intentionally omitted. These optimizations vary across GPU architectures, CUDA versions, and operating systems. Keeping the reference workflow free from hardware-specific optimizations improves portability, simplifies debugging, and provides a consistent baseline for future experimentation.

Models Used

Model

• ltx23DISTILLEDGGUF_q80.gguf

Latent Upscale Model

• Not used in this reference workflow.

  I intentionally omitted latent upscaling to keep the workflow simple and to explore Video-to-Video (V2V) upscaling as a dedicated post-processing workflow. A separate V2V reference workflow is planned.

Text Encoders

• gemma_3_12B_it_fp4_mixed.safetensors

• ltx_2.3_text_projection_bf16.safetensors

VAE

• LTX23_audio_vae_bf16

• LTX23_video_vae_bf16

LoRA

• None

Notes

• Update the resolution, frame rate, clip duration, and prompts to match your project requirements.

• This workflow has been tuned and validated using the LTX 2.3 Distilled GGUF model.

• Replacing the model with a non-distilled (Dev) version may require retuning parameters such as:

  • Sigma schedule

  • Scheduler

  • Sampling steps

  • CFG / Guidance

• If your hardware allows, you can incorporate a latent upscaling stage into the generation pipeline or perform a dedicated Video-to-Video (V2V) upscaling pass after generation.

Why Tiled VAE Decode?

Native Video VAE Decode can exceed the VRAM budget on 8 GB GPUs, resulting in an out-of-memory error during decoding.

This workflow uses Tiled VAE Decode directly to provide more consistent and reliable execution on lower-VRAM hardware.

Reference Pipeline Design

This workflow intentionally generates video directly at the target resolution instead of using the common 0.5× latent generation → 2× latent upscale pipeline.

While the 0.5× approach can reduce generation time and VRAM requirements, it also introduces an additional optimization stage and associated tuning decisions.

The goal of this workflow is to demonstrate the core LTX 2.3 generation pipeline with minimal architectural complexity while providing a stable, reproducible baseline.

To keep the workflow focused and easy to understand, additional refinement passes, latent upscaling, and advanced enhancement stages have been intentionally omitted.

If you would like to incorporate latent upscaling into this workflow, refer to FisherKing-LTX2.3-[GGUF-Distilled]-T2V-ReferenceWorkflow-v1.0, which demonstrates how to integrate a latent upscale model into the generation pipeline.

Alternatively, you can perform a dedicated Video-to-Video (V2V) upscaling pass as a post-processing workflow.

Credits

Thanks to the developers and researchers behind LTX, ComfyUI, GGUF support, and the many open-source projects that made this workflow possible.