Exercise 2: Why Both the Negative Prompt and the Term Weighting Were Needed — Possible Solution ==================================================================== WHAT PASS 3'S NEGATIVE PROMPT ACTUALLY DOES ------------------------------ Per this chapter, Pass 3 adds "deformed hands" (among other terms) to the negative prompt. Per imgai1-5's own CFG formula, this negative embedding replaces the unconditional baseline, so at every denoising step the process is actively steered AWAY from the region of embedding space associated with "deformed hands" specifically — a general, after-the-fact exclusion pressure applied uniformly across however the hands happen to be posed. WHAT PASS 4'S TERM WEIGHTING ACTUALLY DOES ------------------------------ Per this chapter, Pass 4 instead changes the POSE ITSELF that's being requested — "(hands resting on the counter, not visibly gripping anything:1.2)" — and per this chapter's own reasoning, "this doesn't add an anatomical rule the model never learned... it does, however, steer the pose itself toward a simpler, less articulated hand position, which reduces how often the fuzziest, highest-variability poses... get generated in the first place." This works upstream of the negative prompt: it changes what kind of hand configuration is even being requested, rather than only excluding bad outcomes after the fact. WHY NEITHER ALONE IS SUFFICIENT ------------------------------ The negative prompt alone (Pass 3) only excludes a labeled category of bad result ("deformed hands") — it does nothing to reduce how often a genuinely difficult, high-variability pose (per imgai1-8's own explanation of why hands are hard) gets requested in the first place, so the underlying difficulty of the pose remains unchanged, and "deformed" results can still occur in forms not neatly captured by the negative prompt's own wording. Term weighting alone (Pass 4), without the negative prompt, would steer toward a simpler pose but wouldn't provide any active exclusion pressure against whatever residual artifacts still occur even in a simpler pose — there would be nothing in the formula actively steering away from a bad result once one started to form. WHY THE TWO TOGETHER ADDRESS DIFFERENT PARTS OF THE SAME PROBLEM ------------------------------ The negative prompt operates on the OUTPUT side (excluding a known bad category, applied at every step via imgai1-5's own CFG mechanism), while the weighted pose clause operates on the REQUEST side (asking for an inherently less error-prone pose in the first place, per imgai1-7's own weighting mechanism). Combining a request less likely to produce an error with active steering away from the specific error category addresses both halves of the problem this chapter's own Pass 3 first surfaced, which is exactly why the capstone applies them as two separate, sequential passes rather than treating either as sufficient alone. WHY THIS WORKS AS AN ANSWER ------------------------------ It explains precisely what mechanism each pass relies on (imgai1-5's CFG-based exclusion vs. imgai1-7's weighted-request steering), and shows they operate on genuinely different parts of the problem (output exclusion vs. request difficulty), which is why the chapter needed both rather than either alone.