Pakkapon Phongthawee

With GPU, it should be a lot faster, but with one caveat: DiffusionLight currently requires 16GB of VRAM. I'm using RTX3090 / 3090Ti, which takes 30 minutes for a single...

You can use a [Colab version](https://colab.research.google.com/drive/15pC4qb9mEtRYsW3utXkk-jnaeVxUy-0S?usp=sharing&sandboxMode=true) for free. but the result might degrade from the full algorithm. Note that: on Colab you can set EV to `0, -2.5, -5` and...

Our insight right now is the higher, the better. But the score report in the paper we are using - ball_per_iteration: 30 - num_iteration: 2 From our insight, going beyond...

With this paper (DiffusionLight) No. However, with the current state technology, [SDXL Turbo](https://stability.ai/news/stability-ai-sdxl-turbo) mentioned that they can archive 67ms. So, it is possible to port DiffusionLight from [SDXL](https://huggingface.co/docs/diffusers/en/using-diffusers/sdxl) to [SDXL...

the LoRA is only trained to operate only on 1024x1024. You need to retrain LoRA to be able to run on 512x512

As you menntion that this problem is coming from the model being limited by EV-5. To train the model by yourself to support lower than EV-5, we provide a training...

So, you solved the problem by putting image inside the src folder. is that correct?

I tested on Collab, and it doesn't have any errors. Does it have an error on the previous block? from the error message, it seems to have an error from...

It needs an OpenEXR to run. I believe Conda should handle OpenEXR automatically. Please follow our [Getting started](https://github.com/DiffusionLight/DiffusionLight/?tab=readme-ov-file#getting-started) -------------- But in case you are really out of luck with Conda....

@Damojan Please try `pipwin` or `conda-forge` that mention in the above comment If you still cannot make it work, here is your solution ### 1. Using Linux This repo is...