🖼️ LTX2.3 Video Generation using ComfyUI

💀 Why you should use my workflow instead of official workflow ?

This workflow is preferred over the original because the latter imposes strict input constraints, requiring the image width and height to be integers divisible by 64. Additionally, the frame count must be a multiple of 8 plus 1 (e.g., $8n + 1$). In contrast, this updated workflow features an automated adjustment mechanism that seamlessly handles any input dimensions.

💀 Download Workflow

I provide 2 links provider, 1 is my selfhost with faster download speed than GoogleDrive. But GoogleDrive make sure the persistent.

Full mode: ChenjiaDrive / GoogleDrive
Distilled: ChenjiaDrive / GoogleDrive
Kijai version: ChenjiaDrive / GoogleDrive
Preview version: ChenjiaDrive / GoogleDrive

⚠️⚠️⚠️ ⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️

Important !

Do NOT install package ComfyUI-LTXVideo package from official page because it depend on deprecated package.
And cannot be loaded on newest ComfyUI version.
Please install my fork: sandichhuu/ComfyUI-LTXVideo ⚠️⚠️⚠️ ⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️

💀 Hardware Tested

CPU: Ryzen 5 5500.
GPU: RTX 3060 12GB VRAM (1x)
RAM: DDR4 3200 32GB.

Result: 83-578 seconds for video 5 seconds.

💀 Know Bugs

Workflow 1, 2 not working now because the bug from ComfyUI, output is black or errors. You can download first, use later. Workflow 3, 4 using Kijai models and working, but slower than models from Unsloth.

💀 Sorry

I cannot provide preview version using taeltx2_3.safetensor for preview. The preview version using Kijai (text_encoder & vae) is fast enough, only 83 seconds to generate 5 seconds video.