Researchers investigated the role of image tokenizers as "visual languages" within unified multimodal models by developing a controlled, pure-autoregressive testbed. Rather than relying on isolated metrics, the study tracks task-specific validation losses during multimodal continual pretraining across four distinct modalities: text, image, text-to-image (T2I), and image-to-text (I2T) prediction. The findings reveal that loss metrics exhibit divergent scaling behaviors depending on the task, meaning they rank different tokenizers inconsistently if analyzed globally. Crucially, the relationship between loss and downstream performance is highly dependent on the predicted token space; while T2I and I2T losses correlate with generation quality for a fixed tokenizer, the T2I loss-performance relationship shifts when comparing different tokenizers. In contrast, I2T loss—calculated over a shared text vocabulary—provides a more consistent signal and correlates with both generation and visual understanding performance following supervised finetuning. The study demonstrates that superior reconstruction quality does not inherently guarantee lower task-specific losses or improved downstream performance. Furthermore, the choice of image tokenizer significantly impacts text modeling during joint optimization. By examining design axes such as discriminator architecture, semantic supervision, and vocabulary size, the authors highlight the complex interplay between visual and textual tokens in multimodal learning.