Best way to establish a DiffusionLight-Turbo baseline?

[IanMaquignaz]

Hi,

I would like to compare my model to your DiffusionLight-Turbo, but my model is trained against a different HDR dataset.

I assume it would be invalid to collect results without re-training your Lora on my dataset?

Thanks,

[pureexe]

Thank you for your interest in our work.

Regarding your question, DiffusionLight-Turbo is designed for in-the-wild lighting estimation. This means the model is not tailored to any specific dataset, but instead aims to generalize across diverse real-world scenarios. Therefore, it is not necessary to retrain our LoRA on your dataset for a fair comparison.

In addition, a complementary evaluation you may consider is testing your model on the same test set as us. Please refer to the DiffusionLight-Evaluation repository. If your method performs well on the same test set, it would provide stronger evidence that your results are not biased or due to dataset-specific selection.