A question

[xingy038]

Is it possible that the model inputs RGB and the direct outputs depth and normal by modify output layers?

[haodong2000]

Hi @xingy038 thanks for your interest!

We didn't modify any layers in the base model (i.e., w/o LCM) of Lotus-2's core predictor, in order for better preserve the prior knowledge. Other parts of Lotus-2 (i.e., LCM in the core predictor and the detail sharpener) are newly added.

I think modifying some layers is feasible, but it's better to analyze the weights of which layers should be more necessarily maintained at first. For instance, comparing the fine-tuned weights (the base model of Lotus-2's core predictor) and the pre-trained weights of FLUX.