This is a cool and rare occurrence of a non foundation model (94M params). They say, “model still needs work but we would love to start a conversation and gather feedback.”
If I were to postulate, they do note that they have a commercial model (BEN+Refiner), which does even better. So while there is some commercial benefit, the model is released openly, and it’s kind of a win-win situation—openly sharing the model for feedback to use, and then using said feedback to improve both the base model and the commercial model.
One feedback is that the output is blurry on their example image, and the OP follows up by how it was trained on 1024x1024 images, and that they will be working on a dynamic gaussian blur algorithm to mitigate the effects.
Another feedback is to add weights_only=True to the torch.load() function call, which avoids the execution of arbitrary code.
Another asks integration with comfyUI, and OP announces that it will be supported in the coming days.
Another asks for a HF spaces demo.
One thing I am noticing (and potentially will notice more of in the future) is that most of the feedback are to do with usability, like good integration, and how accessible it is (comfyUI, HF spaces).