LTX-2.3-RTX5060Ti

<i>Let it be known that I am no longer supporting LTX and have moved on to MiniMax H3. This will be my first and last workflow for LTX as I have attempted to cleanup LTX defects to the best of my ability, but audio dialogue still remains to be a problem that will never be resolved. Honestly, dump LTX and go to H3. The generation times are longer, but it has much better prompt adherence.</i>

<b>Introduction</B>:

This ComfyUI workflow was designed to correct the flaws of other workflows published on the Internet, eliminating problematic 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 720p. Believe it or not, this workflow can create near perfect 60 second videos with minimal artifacting or warble effects. 30 second videos are completed in roughly 10 minutes as the quality improves per pass. Naturally if you use the Omni-RL-NFT LoRA, that's going to increase render times.

<B>RAM Requirements</B>:

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

<B>How-To-Resolution:</B>

If you start bleeding over 300 MB (0.3GB) into system RAM, back down your resolution. Going over 0.3GB will increase your render times dramatically. People also tend to get their resolution settings wrong. LTX 2.3 requires 'divisible by 32' resolutions. The supported list is the following:

• 640x384
• 768x448
• 896x512
• 1024x608
• 1280x704 <- This is where you want to be for 720p
• 1600x896
• 1920x1088 <- This is where you want to be for 1080p
• 2048x1152

If you want 9:16, swap the X and Y resolutions to match.

<B>Seconds to Frames Converter For Frame Accuracy</B>:

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

<B>Required nodes</B>:

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

<B>Recommended Samplers and Schedulers:</B>

• Res Multistep (Res_2S) / Simple - Quick renders in high quality with 4 to 6 steps when using a distilled model.
• DPMPP 2M / Simple
• Euler Ancestral Config Post Processing / Simple

<B>Additionals</B>:

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.