Understanding the Impact of Separated Image Encoder on Training and Image Generation Quality

[lavinal712]

Hello,

I am currently exploring the impact of separating the image encoder in our model architecture on both the training process and the quality of generated images. Specifically, I am interested in understanding whether this separation helps or hinders our ability to understand the images better. Additionally, I would like to discuss the potential relationship between the two aspects—whether they are mutually beneficial or if one negatively affects the other.

Key Questions:
Separation Impact: Does separating the image encoder from the rest of the model architecture help in better understanding the images during training?

Mutual Relationship: Are the training process and the quality of generated images mutually beneficial, or does one negatively impact the other?

Janus and Janusflow Quality: After conducting experiments, we observed that the quality of images generated by Janus and Janusflow is not satisfactory. Does this indicate that the method is not feasible, or are there potential improvements or adjustments that could enhance the results?