[Feature Request] The ability to use TensorRT Models, includes evaluation

[CloudWalker-II]

I would love to be able to use TensorRT-trained models within Krita,

https://github.com/comfyanonymous/ComfyUI_TensorRT

For those who may not be familiar, TensorRT allows models on Nvidia GPUs to leverage the RT cores, resulting in a significant boost in performance.

I think this could be particularly useful for the live mode feature. While TensorRT models don’t currently support adding Loras after training, the speed improvements are impressive—comparable to using the Hyper model settings, in my experience, but without any noticeable loss in quality. You could even train a model with an already integrated Hyper Lora (or lighting, turbo, etc) to double down on the speed gains from both.

Here is some comparisons from my own tests. The results are from a rtx 3080 10GB VRAM, I used a SDXL model at 1024 x 1024px and done with 25 steps, no Hyper-Lora. All images between the normal model and the RT enabled version where visually identical.

TestPrompt 1
Euler_a, Simple Sampler
Normal Model	3.05it/s, Prompt executed in 11.11 seconds
RT Engine	4.90it/s, Prompt executed in 5.53 seconds
Dpmpp_2m_sde, Karras
Normal Model	2.96it/s, Prompt executed in 10.41 seconds
RT Engine	4.64it/s, Prompt executed in 5.79 seconds

TestPrompt 2
Euler_a, Simple Sampler
Normal Model	3.12it/s, Prompt executed in 10.20 seconds
RT Engine	4.86it/s, Prompt executed in 5.55 seconds
Dpmpp_2m_sde, Karras
Normal Model	2.94it/s, Prompt executed in 10.27 seconds
RT Engine	4.55it/s, Prompt executed in 5.90 seconds