Exercise 1: Why Term Weighting Is a More Granular Version of CFG Scale — Possible Solution ==================================================================== WHAT CFG SCALE CONTROLS, PER imgai1-5 ------------------------------ Per imgai1-5's own formula, CFG scale amplifies the difference between the conditional prediction (steered by the WHOLE prompt's text embedding) and the unconditional prediction, at every denoising step. It's a single dial applied uniformly to the entire prompt's own overall pull — there's no way, with CFG scale alone, to say "amplify THIS part of the prompt more than THAT part." WHAT TERM WEIGHTING ADDS ------------------------------ Per this chapter, term weighting lets you control "how strongly one specific word or phrase contributes to that overall pull, relative to the rest of the prompt" — for example, Stable Diffusion's "(red umbrella:1.4), rainy street" or Midjourney's "red umbrella::2 rainy street::1". Rather than amplifying the entire prompt's own conditioning signal uniformly (what CFG scale does), term weighting amplifies or reduces the contribution of a SPECIFIC sub-part of that same conditioning signal, before the whole thing gets combined and (in Stable Diffusion's case) run through the same CFG amplification imgai1-5 already described. WHY "MORE GRANULAR VERSION," NOT "UNRELATED TECHNIQUE" ------------------------------ Both techniques do the exact same underlying kind of thing — amplifying how strongly some piece of text-derived conditioning signal pulls the denoising process toward it — they just operate at different scopes. CFG scale operates at the scope of "the entire prompt vs. no prompt at all." Term weighting operates at the scope of "this one phrase vs. the rest of the same prompt." One is the coarse, whole- prompt version of exactly the same underlying lever the other applies at a finer, per-term grain — which is precisely what makes "more granular version of the same idea" (this chapter's own framing) an accurate description rather than a loose analogy. WHY THIS CONNECTS TO THE CHAPTER'S OWN WARN-BOX ------------------------------ Because both techniques are the same underlying kind of amplification lever, both fail in the same way when pushed too far — per this chapter's own warn-box, "a term pushed to an extreme weight distorts the same way an overall CFG scale pushed too high does," which only makes sense if the two really are variations on one mechanism rather than two unrelated ones. WHY THIS WORKS AS AN ANSWER ------------------------------ It identifies the shared underlying mechanism (amplifying a text- derived conditioning signal's pull) both techniques rely on, explains precisely what scope each one operates at, and uses the chapter's own warn-box (both fail identically when pushed too far) as confirming evidence that they're the same lever at different granularities.