Increase Trellis realism ? combat cartoonishing !

Hello,

First, I would like to express my amazement on this impressive work. I just have noticed that sometimes Trellis tends to generate cartoon-like assets, perhaps because such examples are present in the dataset.

I had the idea of trying to address this by encouraging the generation of more realistic assets. Could the authors (@YuDeng, @JeffreyXiang, @yangjiaolong) suggest possible directions I could explore? I have read the paper carefully, and although I am only a Master’s student in AI with humble knowledge of deep learning, I am eager to learn and contribute. For instance, would fine-tuning the Flow-Transformer be a reasonable approach?

I would be very grateful for any guidance, including papers or methods I could study to better tackle this problem.

cartoonished assets :

No repository files, tests, or entry points are identified. Start by reviewing the paper and the Flow-Transformer approach mentioned in the issue, then define a concrete change and evaluation for reducing cartoon-like assets. The work is done only when its scope and measurable outcome are established.