Pakkapon Phongthawee

I don't directly use Triton. I don't know which library uses Triton and how it will affect the inference speed. However, 20 min per input image is expected for A6000...

We don't have and didn't plan for a proper multi-GPU support. However, if you have multi GPU. Here is the method that can help you get the chromeball faster. ##...

This happened because we assumed [orthographic projection](https://en.wikipedia.org/wiki/Orthographic_projection) to unwrap the chromeball, which is not true in reality. Given our setup (using a single image, unknown camera calibration), I don't think...

With orthographic assumption, we assume that chromeball covers $\[0,2\pi\]$ in theta (horizontal) and $\[0,\pi\]$ in phi (vertical), which leads to the artifact. https://github.com/DiffusionLight/DiffusionLight/blob/3e90bfc7eda193709cc4bc0c35acc59c08446d70/ball2envmap.py#L35-L36 But in reality, we need to know...

I want to try it out but I have the same problem. ROG Phone 6 running Android 13

DiffusionLight requires a GPU with at least 16GB of VRAM, as inference typically consumes around 13–14GB of VRAM. Your RTX 3070 laptop, with only 8GB of VRAM, is likely running...

Sadly, our method doesn't have temporal consistency. The chromeball will be different on each generation. I suggest you generate only single chromeball per video using the first frame or middle...

Our implementation typically consumes around 16GB of VRAM per image, so it's not feasible to handle multiple requests simultaneously on a single GPU due to memory limitations. As an alternative,...

We used blender to render 3D model with the environment map extract from DiffusionLight We have a tutorial on [repo's wiki](https://github.com/DiffusionLight/DiffusionLight/wiki/How-to-Object-insertion-using-DiffusionLight!)

Let's seperate the question into multiple question ## Can we input 4K images and receive 4K output? Technically, you can. But you shouldn't. Because [SDXL](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) that DiffusionLight based on is...