The Mid-Noise Paradox

Senior ML Engineer interview at Midjourney, and the interviewer asks:

“Your DiT trains beautifully at 512×512. You bump inference to 1024×1024 and it generates garbage, warped anatomy, repeated limbs, a teddy bear with three faces. Before you touch the VAE or the sampler, where do you look first?”

Don’t say: “The model didn’t see high-res data, so I’ll fine-tune at 1024.”