Worked Example: Sampling a 2048x2048 Image

Your image generator was trained with a tokenizer where each visual token corresponds to a 64x64 pixel chunk. The product now needs 2048x2048 output, but you'd rather not retrain the generator at that resolution.

1. If you generated natively at 2048x2048 with 64px tokens, how many sequential generation steps would that take, and why is that a latency concern worth flagging?
2. Propose an alternative that avoids retraining the generator at the higher resolution, and name the general pattern it's an instance of.
3. Walk through the full sampling process (seed to final pixels) for your proposed approach, including where top-p sampling fits in and what it's trading off.