Exercise 1: Why CFG Scale 1 Behaves Like the Conditional Prediction Alone, and Why High Values Overcook — Possible Solution
====================================================================

THE FORMULA ITSELF
------------------------------
Per this chapter:
  final_prediction = unconditional_prediction
                    + CFG_scale × (conditional_prediction − unconditional_prediction)

WHY CFG SCALE 1 REDUCES TO ROUGHLY THE CONDITIONAL PREDICTION
------------------------------
Substituting CFG_scale = 1 into the formula:
  final_prediction = unconditional_prediction + 1 × (conditional_prediction − unconditional_prediction)
                    = unconditional_prediction + conditional_prediction − unconditional_prediction
                    = conditional_prediction

The unconditional_prediction terms cancel out algebraically, leaving
just the conditional_prediction -- exactly matching this chapter's own
claim that "a CFG scale of 1 uses the conditional prediction roughly
as-is." This isn't a coincidence or an approximation; it falls directly
out of the formula's own arithmetic.

WHY HIGHER VALUES AMPLIFY THE PROMPT'S PULL
------------------------------
Per this chapter, "higher values... amplify the direction the text
embedding is already pulling toward, producing results that adhere to
the prompt more strongly." Values above 1 multiply the
(conditional_prediction − unconditional_prediction) difference by more
than its own natural size, pushing the final prediction further in
that same direction than the conditional prediction alone would go --
essentially exaggerating however the text prompt was already nudging
the image, beyond what a single, un-amplified conditional prediction
would produce.

WHY THIS OVERSHOOTS RATHER THAN BECOMING "MORE ACCURATE"
------------------------------
Per this chapter, "pushed too high, that amplification overshoots --
oversaturated colors, harsh contrast, and visible artifacting." The
(conditional_prediction − unconditional_prediction) difference at any
given step is only ever a REASONABLE nudge in the direction the text
prompt is pulling -- it's not itself a value the model was ever
trained to receive amplified by an arbitrarily large multiplier. Once
CFG_scale grows large enough, the formula pushes the prediction well
outside the range of values the denoising network was actually trained
to interpret sensibly, in the same direction repeatedly and
compoundingly across many steps -- producing exactly the kind of
distortion (oversaturation, harsh contrast, artifacting) that comes
from over-amplifying a genuinely correct-but-modest signal rather than
from the model becoming somehow more "accurate" to the prompt's own
intent.

WHY THIS WORKS AS AN ANSWER
------------------------------
It shows algebraically, using the chapter's own formula, exactly why
CFG_scale = 1 cancels out to the conditional prediction alone, and
explains mechanically -- via repeated over-amplification of a modest
per-step nudge -- why high values produce distortion rather than
improved accuracy.