<h1 id="workflow-goal">Workflow Goal</h1><p>Provide a clean, educational reference implementation for <strong>LTX 2.3 Text-to-Video</strong> generation.</p><p>This workflow intentionally prioritizes <strong>readability, reproducibility, and educational value</strong> over maximum performance or feature completeness. It serves as a stable reference implementation that users can understand before adding their own optimizations.</p><p>Treat this workflow as a starting point, then customize the prompts and generation settings to match your preferred artistic direction.</p><hr /><h1 id="version">Version</h1><p><strong>v1.0</strong></p><hr /><h1 id="purpose">Purpose</h1><ul><li><p>✅ Educational Reference Workflow</p></li><li><p>✅ Clean Baseline Implementation</p></li><li><p>✅ Dynamic VRAM Friendly</p></li><li><p>✅ Tiled VAE Decode</p></li></ul><hr /><h1 id="features">Features</h1><ul><li><p>✅ Wireless Modular Architecture</p></li><li><p>✅ Clean Reference Implementation</p></li><li><p>✅ Prompt Engineering Ready</p></li><li><p>✅ Two-Phase Sampling Pipeline</p></li><li><p>✅ Low VRAM Compatible</p></li><li><p>✅ Modular and Easily Extensible</p></li></ul><hr /><h1 id="workflow-pipeline">Workflow Pipeline</h1><pre><code>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
</code></pre><hr /><h1 id="verified-settings">Verified Settings</h1><p>The following settings were used to validate this workflow and are recommended as the baseline configuration.</p><h2 id="sampling">Sampling</h2><ul><li><p><strong>CFG:</strong> 1.0</p></li><li><p><strong>Steps:</strong> 11 (8 + 3)</p></li><li><p><strong>Sampler:</strong> Euler Ancestral</p></li></ul><h2 id="video">Video</h2><ul><li><p><strong>Frames:</strong> 145 (≈6 seconds @ 24 FPS)</p></li><li><p><strong>Output FPS:</strong> 24</p></li><li><p><strong>Tested Resolution:</strong> 1024 × 576 (16:9)</p></li></ul><hr /><h1 id="required-lora-configuration">Required LoRA Configuration</h1><p>No LoRAs were used during validation.</p><p>This workflow has been validated using the <strong>LTX 2.3 Distilled GGUF</strong> model. The official Distilled LoRA is <strong>not required</strong> when using the distilled model and has therefore been omitted from this reference workflow.</p><hr /><h1 id="hardware-and-resolution-notes">Hardware &amp; Resolution Notes</h1><h2 id="validated-on">Validated On</h2><h3 id="hardware">Hardware</h3><ul><li><p><strong>GPU:</strong> NVIDIA RTX 2080 (8 GB VRAM)</p></li><li><p><strong>System RAM:</strong> 64 GB</p></li></ul><h3 id="software">Software</h3><ul><li><p><strong>ComfyUI:</strong> 0.27.x</p></li><li><p><strong>Python:</strong> 3.13.7</p></li><li><p><strong>PyTorch:</strong> 2.12.0 + CUDA 13.0</p></li></ul><h3 id="validation-settings">Validation Settings</h3><ul><li><p><strong>Resolution:</strong> 1024 × 576</p></li><li><p><strong>Length:</strong> 6 Seconds</p></li><li><p><strong>Generation Time:</strong> ~12–14 Minutes</p></li></ul><p>Although compatible with <strong>8 GB VRAM</strong>, LTX 2.3 Text-to-Video remains computationally intensive and memory hungry.</p><p>System RAM plays an equally important role in handling intermediate tensors and memory paging during video generation.</p><h2 id="expected-behavior">Expected Behavior</h2><h3 id="8-gb-vram-+-16-gb-ram">8 GB VRAM + 16 GB RAM</h3><ul><li><p>Possible out-of-memory errors</p></li><li><p>Slower generation</p></li></ul><h3 id="8-gb-vram-+-32-gb-ram">8 GB VRAM + 32 GB RAM</h3><ul><li><p>Better stability</p></li><li><p>Performance depends on available system memory</p></li></ul><h3 id="8-gb-vram-+-64-gb-ram-(or-more)">8 GB VRAM + 64 GB RAM (or more)</h3><ul><li><p>Recommended configuration</p></li></ul><p>If your system has a more capable GPU, consider increasing the output resolution to <strong>720p or higher</strong>.</p><hr /><h1 id="design-philosophy">Design Philosophy</h1><p>This workflow intentionally avoids unnecessary complexity.</p><p>The objective is to provide a stable, reproducible reference implementation that users can understand, learn from, and extend for their own creative projects.</p><p>Advanced optimizations such as <strong>SageAttention, Flash Attention, or other hardware-specific accelerations</strong> 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.</p><hr /><h1 id="models-used">Models Used</h1><h2 id="model">Model</h2><ul><li><p><strong>ltx23DISTILLEDGGUF_q80.gguf</strong></p></li></ul><h2 id="latent-upscale-model">Latent Upscale Model</h2><ul><li><p>Not used in this reference workflow.</p><p>I intentionally omitted latent upscaling to keep the workflow simple and to explore <strong>Video-to-Video (V2V)</strong> upscaling as a dedicated post-processing workflow. A separate V2V reference workflow is planned.</p></li></ul><h2 id="text-encoders">Text Encoders</h2><ul><li><p><strong>gemma_3_12B_it_fp4_mixed.safetensors</strong></p></li><li><p><strong>ltx_2.3_text_projection_bf16.safetensors</strong></p></li></ul><h2 id="vae">VAE</h2><ul><li><p><strong>LTX23_audio_vae_bf16</strong></p></li><li><p><strong>LTX23_video_vae_bf16</strong></p></li></ul><h2 id="lora">LoRA</h2><ul><li><p>None</p></li></ul><hr /><h1 id="notes">Notes</h1><ul><li><p>Update the resolution, frame rate, clip duration, and prompts to match your project requirements.</p></li><li><p>This workflow has been tuned and validated using the <strong>LTX 2.3 Distilled GGUF</strong> model.</p></li><li><p>Replacing the model with a non-distilled (<strong>Dev</strong>) version may require retuning parameters such as:</p><ul><li><p>Sigma schedule</p></li><li><p>Scheduler</p></li><li><p>Sampling steps</p></li><li><p>CFG / Guidance</p></li></ul></li><li><p>If your hardware allows, you can incorporate a latent upscaling stage into the generation pipeline or perform a dedicated <strong>Video-to-Video (V2V)</strong> upscaling pass after generation.</p></li></ul><hr /><h1 id="why-tiled-vae-decode">Why Tiled VAE Decode?</h1><p>Native <strong>Video VAE Decode</strong> can exceed the VRAM budget on <strong>8 GB GPUs</strong>, resulting in an out-of-memory error during decoding.</p><p>This workflow uses <strong>Tiled VAE Decode</strong> directly to provide more consistent and reliable execution on lower-VRAM hardware.</p><hr /><h1 id="reference-pipeline-design">Reference Pipeline Design</h1><p>This workflow intentionally generates video directly at the target resolution instead of using the common <strong>0.5× latent generation → 2× latent upscale</strong> pipeline.</p><p>While the 0.5× approach can reduce generation time and VRAM requirements, it also introduces an additional optimization stage and associated tuning decisions.</p><p>The goal of this workflow is to demonstrate the core <strong>LTX 2.3</strong> generation pipeline with minimal architectural complexity while providing a stable, reproducible baseline.</p><p>To keep the workflow focused and easy to understand, additional refinement passes, latent upscaling, and advanced enhancement stages have been intentionally omitted.</p><p>If you would like to incorporate latent upscaling into this workflow, refer to <strong>FisherKing-LTX2.3-[GGUF-Distilled]-T2V-ReferenceWorkflow-v1.0</strong>, which demonstrates how to integrate a latent upscale model into the generation pipeline.</p><p>Alternatively, you can perform a dedicated <strong>Video-to-Video (V2V)</strong> upscaling pass as a post-processing workflow.</p><hr /><h1 id="credits">Credits</h1><p>Thanks to the developers and researchers behind <strong>LTX</strong>, <strong>ComfyUI</strong>, <strong>GGUF support</strong>, and the many open-source projects that made this workflow possible.</p>

Files with SHA256: cc2e6de95446c6e524c8eb2e9a2edd17806a592349965524169163076a43353a