Diagnosing a Blurry VQGAN Reconstruction

Your team trains an image tokenizer using only reconstruction loss (pixel MSE) and quantization loss. At low resolution the reconstructions look fine, but at 1024x1024 the decoder output is noticeably blurry and loses fine texture (skin pores, fabric weave, individual leaves), even though the overall shapes and colors are correct.

1. Explain why reconstruction + quantization loss alone produces this specific failure mode (blurry but structurally correct).
2. Which two additional losses would you add, and what does each fix that the other doesn't?
3. After adding both, how would you know the fix actually worked, beyond "the images look nicer to me"?