Karan Singh

Hey @OPHoperHPO @senya-ashukha , First of all thanks for this conversion, I was trying the ONNX model, and it seems like it differs from the original model and doesn't perform...

Thanks for the clarification @OPHoperHPO, I used the same preprocessing/postprocessing functions as provided in your conversion [notebook ](https://colab.research.google.com/github/Carve-Photos/lama/blob/main/export_LaMa_to_onnx.ipynb#scrollTo=AEcIuUpleXuZ) for the onnx inference, as for the pytorch inference, I followed [Sanster's](https://github.com/Sanster)...

@OPHoperHPO, I have created an isolated [notebook](https://colab.research.google.com/drive/19uZvT8HnTfXlk_lSAWUhJA9WJ2I6JXPc?usp=sharing) and removed all the external factors(courtesy [here](https://github.com/Sanster/IOPaint/blob/main/iopaint/model/lama.py)). I also tried replacing just the inference part via onnx and keeping the same preprocessing logic...

@OPHoperHPO Haah, thanks catching my little blunder ;-; The onnx model is indeed giving similar outputs 🙌 Big big thanks for your help!

@ljdang fp16 doesn't seem to have any visible losses, is your conversion correct?

any update on this, after the PowerPaint_v2 some conversion guidelines would be really useful

Seems like it stopped at: Start to Optimize the MNN Net...  This didn't throw any error, also the ONNX model runs properly.

any information on this @jxt1234

@k-khosravi can you provide the code you used for onnx conversion?

@jxt1234 we would love to chat about this and make [mnn-rs](https://github.com/aftershootco/mnn-rs) more usable, among the rust community!