Smite79/LTX-2.3-RTX5060Ti

This ComfyUI workflow was designed to correct the flaws of other workflows published on the Internet, eliminating outdated KJNodes and reducing the process to a single flow. It was also designed for users with mid-tier GPU's like the RTX 5060Ti, allowing you to produce high quality videos with the mxfp8 model while remaining in the VRAM budget for 16GB models with minimal leaks to system RAM. Max supported resolution is 1024x576. Believe it or not, this workflow can create near perfect 60 second videos with minimal artifacting or warble effects.

The minimum RAM requirement for this workflow is 64GB. Running this workflow will consume 36-40GB, depending the LoRA's you choose to use.

Drop my .py file into your custom_nodes folder. It is used to convert seconds to frames using 8n+1.

Required nodes: DaSiWa's for LoRA loading, Winnougan's for taeltx-preview and NAG guidance. SeanScripts' Unload Model node for RAM purging after completion.

For best experience, ensure that you have Sage Attention 2.2.0 installed with the CUDA 13.3 toolkit. It is fully compatible with UNet's, NVFP4, BF16 and FP8 models.

To keep RAM usage contained, use the following launcher flags: --use-sage-attention --enable-triton-backend --auto-launch --enable-dynamic-vram --disable-smart-memory --disable-pinned-memory --fast-disk --fp16-intermediates

You choose to use this workflow at your own risk. I am not providing technical support for it. This workflow is not recommended for use with Ampere GPU's.