ComfyUI: Seed hunting with a multi-stage LTX 2.3 workflow
ComfyUIDemo summary
The creator demonstrates his custom ComfyUI workflow that generates four low-resolution LTX 2.3 samples simultaneously to find a 'golden seed' before upscaling to 1080p.
Step-by-step
- Run the first stage to generate four low-resolution samples with unique seeds
- Review the four sample outcomes to identify a 'golden seed'
- Toggle Finish Mode to enable the second and third stages
- Select the preferred sample latent number to proceed with finalization
- Monitor the second stage as it fine-tunes visuals and motion against the reference image
- Allow the workflow to proceed to the third stage for 1080p upscaling
Options
- Cancel a specific stage if you don't like the result
- Click the reroll button on a specific stage to redo it without restarting the entire workflow
Tips
- Use the low-resolution first stage to save time, as it typically takes only 20 to 60 seconds to find a good seed
Highlights
“This process is similar to gold panning”
All demos from “AI Filmmaking Part 1 | Video gen with LTX 2.3: You're doing it wrong.”
- 1:420:59Seed hunting with a multi-stage LTX 2.3 workflowCurrentThe creator demonstrates his custom ComfyUI workflow that generates four low-resolution LTX 2.3 samples simultaneously to find a 'golden seed' before upscaling to 1080p.ComfyUI· AI Animation Generator
- 4:251:21Iterative rerolling at specific generation stagesThe video shows how to use the 'reroll' button at specific stages of the workflow to change motion or visual quality without re-generating the initial seed samples.ComfyUI· AI Image Generator
- Watch “AI Filmmaking Part 1 | Video gen with LTX 2.3: You're doing it wrong.” →
AI Animation Generator
2:070:23Load Wan 2.2 Animate workflow and install nodesThe user demonstrates how to drag and drop the Wan 2.2 Animate workflow into ComfyUI and use the Manager to install missing custom nodes.MDMZ- 2:551:05Configure video input and output settingsThe demo shows how to upload a source video to ComfyUI, set the frame count, and adjust the output dimensions to match the original aspect ratio.MDMZ
16:361:56Setting up HunyuanVideo 1.5 in ComfyUIThe video demonstrates how to update ComfyUI and import the HunyuanVideo 1.5 JSON workflow files to create a node-based generation environment.AI Search- 20:151:53Text-to-Video generation in ComfyUIA step-by-step demo of configuring the Hunyuan nodes in ComfyUI, entering a prompt for a 'giant cat', and rendering the final 720p video.AI Search
- 29:151:33Running HunyuanVideo with GGUF (Low VRAM)The video shows how to use the GGUF loader node to run a compressed version of HunyuanVideo 1.5, enabling video generation on GPUs with as little as 6GB of VRAM.AI Search
0:594:52Configure LTX 2.3 in ComfyUIThe creator walks through the ComfyUI node setup for LTX 2.3, explaining the GGUF model loader, VAE settings, and how to adjust resolution and frame counts for optimal rendering.AIKnowledge2Go
1:230:41Setting up LTX-2.3 in ComfyUIThe creator demonstrates how to browse templates in ComfyUI, search for LTX 2.3, and download the required missing models for the text-to-video workflow.MDMZ
13:071:21Applying LoRA to Wan 2.2 video generationThe video shows how to integrate a LoRA (Low-Rank Adaptation) into a Wan 2.2 workflow to achieve specific cinematic movements like a face zoom.pixaroma- 31:133:44Text-to-video with LTX-2The video walks through setting up the LTX-2 model in ComfyUI to generate high-resolution video clips from text prompts and images.pixaroma
- 39:075:27Cloud-based ComfyUI on RunPod/RunHubThe video shows how to run complex video workflows in the cloud using RunHub AI, demonstrating the interface and execution of InfiniteTalk and Wan 2.2 without local hardware.pixaroma
2:351:46Configure Infinite Talk models in ComfyUIThe creator demonstrates how to organize the necessary models within the ComfyUI workflow, including the Lightning LoRA, quantized Infinite Talk UNET models, and the Wan 2.1 VAE and Clip Vision nodes.Aiconomist
11:060:26Combining audio, images, and promptsA demonstration of layering a specific action prompt (patting stomach) over a specific audio timestamp to create a fully directed AI scene.What Dreams Cost