The RTX 4070 performance in LTX 2.3?

[gamevalley123-cloud]

I'm curious about the performance of the RTX 4070 when using LTX 2.3 in ComfyUI for a 10 second long 720 clip (in both T2V and I2V workflows). If anybody could share with me their experience, it will really be very helpful. It'd also be great if you could share with me the workflow, the settings and the model you're using.

[Caacet]

4070 TI Super, 32 GB RAM, GGUF Q5 with distilled lora, some default worklow I2V. (first pass around 800x600, second pass x1.5). hot start - 320-600 sec to full 8sec vid. Pain is random time and "size" or "shards" of rotating models in memory. First Ksampler generates 8 steps - 1.3 min , then next job, all same workflow - 3 min... With default comfy arguments all models offloaded after each job... I tried --disable-dynamic-vram, --disable-smart-memory. I haven't found any normal set of arguments yet, it's just random. But it's obvious that my config can achieve faster performance. The issue is stability; it works, it can work fast but extremely inconsistently.