InfiniteTalk inference is TOO SLOW. How did you speed up inference?

[AhmadAFS1]

Hi all,
I just discovered infiniteTalk recently. So I rented an rtx 4090 24gb and tried to see what results I could get.
I noticed that its pretty slow, so I'm wondering what optimizations you guys did to speed things up.
For reference the quality is amazing, but it takes 40 minutes for a 10 second video at 480p. For this initial test I did not use any quantized models or any loras.

From what I researched, the follow optimizations are what I'd probably need to do, but I want to validate with those more experienced with the software on what the best optimizations actually are which shouldn't degrade quality too much. My initial expectation was being able to achieve a 5 second clip within 5 minutes.

potential optimizations I'd have to do:

• use FusionX and lightx2v Loras
• use teaCache
• disable the low vram mode
• install SageAttention

MY QUESTION: what other optimizations do I need to do in order to get faster generation without sacrificing quality too much? Ideally achieving a 5 second within 5 minutes.

This is what I ran with:
Hardware:

• GPU: RTX 4090 / 24 GB VRAM
• CUDA/PyTorch: torch 2.4.1 + cu121
• Running locally in a conda env

Model/settings:

• Task/model: InfiniteTalk 14B
• Base checkpoint: Wan2.1-I2V-14B-480P
• InfiniteTalk checkpoint: single-person infinitetalk.safetensors
• Resolution preset: infinitetalk-480
• Mode: streaming
• sample_steps: 8
• frame_num: 81
• max_frame_num: 233
• offload_model: True
• num_persistent_param_in_dit: 0
• LoRA: none
• TeaCache: disabled
• SageAttention: not installed
• FlashAttention: installed/active