Exercise 2: Why a Negative Prompt Is a Full Second Prompt, Not a Simple Filter — Possible Solution ==================================================================== THE TWO FORMULAS SIDE BY SIDE ------------------------------ Without a negative prompt, per this chapter: final_prediction = unconditional_prediction + CFG_scale × (conditional_prediction − unconditional_prediction) With a negative prompt, per this chapter: final_prediction = negative_prediction + CFG_scale × (positive_prediction − negative_prediction) WHAT LITERALLY CHANGES BETWEEN THE TWO ------------------------------ Every occurrence of unconditional_prediction in the first formula is replaced by negative_prediction in the second. Per this chapter, "a negative prompt is itself run through CLIP the same way the positive prompt is, and its own embedding replaces the unconditional prediction in the formula." The unconditional_prediction represents "as if no text prompt existed at all" (per this chapter's own CFG section) -- a neutral baseline with no directional pull of its own. negative_prediction, by contrast, is a real, text-conditioned prediction, steered by CLIP toward whatever the negative prompt describes, exactly the same way the positive prediction is steered toward the positive prompt. WHY THIS MEANS ACTIVE STEERING AWAY, NOT JUST EXCLUSION ------------------------------ Per this chapter, "this doesn't just avoid the negative prompt's content -- it actively steers generation AWAY from that region of embedding space, at every single denoising step, amplified by the same CFG scale governing how strongly the positive prompt pulls toward its own target." Because the formula computes (positive_prediction − negative_prediction), and negative_prediction is itself pulled toward the negative prompt's own embedding region, the resulting difference vector points not just "toward positive" in a vacuum, but specifically away from the negative prompt's region at the same time. A simple exclusion filter would only need to suppress or block certain content after generation; this mechanism instead actively redirects every single denoising step's own trajectory before the image is even formed. WHY THIS QUALIFIES AS A "FULL, FIRST-CLASS SECOND PROMPT" ------------------------------ The negative prompt receives the exact same treatment as the positive prompt at every stage: it's run through CLIP to produce a real text embedding (not a keyword list), and that embedding directly and continuously participates in the same per-step amplification mechanism (CFG_scale) that governs the positive prompt's own influence. Nothing about its role in the formula is a lesser, bolted-on, or after-the-fact mechanism -- it occupies the exact same structural position the unconditional baseline used to occupy, just filled with real, directed content instead of neutrality. WHY THIS WORKS AS AN ANSWER ------------------------------ It compares the two formulas term by term, shows precisely which term changes and what it's replaced with, and explains why replacing a neutral baseline with a real, CLIP-embedded prediction makes the negative prompt an active steering mechanism rather than a passive exclusion filter.