Generative Vision Interview Questions #4 - The SNR Collapse Trap

Why skipping continuous float scaling doesn't just cause NaN losses, it mathematically dwarfs your scheduled noise and leaves your reverse process dead on arrival.

You’re in a Senior AI Engineer interview at Midjourney. The interviewer sets a trap:

“Images are discrete RGB values from 0 to 255. Diffusion math assumes we sample from a standard normal distribution. If a data engineer feeds raw 0-255 pixel tensors directly into the training pipeline without continuous float scaling, how does this mathematically break the variance-preserving nature of the forward process?”

90% of candidates walk right into it.

The textbook instinct is to blame the neural network’s mechanics.

Most candidates say: “It will cause exploding gradients. The unscaled inputs will saturate the activation functions, and your loss will immediately spike to NaN.”

While that might be true for the U-Net, it completely misses the mathematical foundation of diffusion.

But you aren’t debugging a standard image classifier; you are debugging a stochastic Markov chain.

The reality is that the forward process q(xₜ|xₜ₋₁) relies on a strictly calibrated noise schedule to incrementally destroy the image. If you don’t scale the inputs to [-1, 1], you trigger what I call The 𝐒𝐍𝐑 𝐂𝐨𝐥𝐥𝐚𝐩𝐬𝐞.

Here is what is actually happening under the hood:

• The Massive Variance: A uniform distribution of pixels from 0 to 255 has a variance of roughly 5,400.

• The Microscopic Noise: A standard variance-preserving schedule injects noise using βₜ values that start infinitesimally small (e.g., 1e-4 at t=1).

• The Math Breakdown: You are adding a tiny fraction of 𝒩(0, I) noise to a signal with massive magnitude. The injected noise becomes a literal rounding error.

• The Convergence Failure: By step T=1000, your noisy image x_T is supposed to equal pure standard Gaussian noise. Unscaled, it never even gets close.

If x_T doesn’t converge to 𝒩(0, I), your reverse process is dead on arrival. You will be asking the model to denoise from a distribution it has never seen.

---

The Answer That Gets You Hired:

“Feeding raw 0-255 inputs breaks the boundary conditions of the Markov chain. The data variance dwarfs the scheduled noise variance, meaning x_T never converges to a standard normal distribution. The model will fail to generate anything because it’s expecting to start denoising from pure 𝒩(0, I), but your forward process never actually reached it.”

---