Exercise 3: Why This Capstone Was Built on Stable Diffusion Specifically — Possible Solution ==================================================================== THE CHAPTER'S OWN OPENING REASONING ------------------------------ Per this chapter's own opening, Stable Diffusion was chosen "specifically because it's the only tool in this course whose mechanism (imgai1-2) can be reasoned about precisely at every step, rather than described only behaviorally (imgai1-4) or mediated through a rewriting layer (imgai1-6)." WHY MIDJOURNEY WOULD HAVE LIMITED THE CAPSTONE'S OWN EXPLANATORY DEPTH ------------------------------ Per imgai1-4's own warn-box, Midjourney's parameters can only be described by their observed, documented behavior — its architecture is never published. A capstone built on Midjourney could still show a working iteration workflow (adjusting --stylize, --chaos, --no), but it could never explain, mechanically, exactly why a given adjustment produces a given change the way this chapter's own Pass 3 could explain the CFG/negative-prompt formula precisely, per imgai1-5. WHY DALL-E WOULD HAVE INTRODUCED A DIFFERENT, MEDIATED PROBLEM ------------------------------ Per imgai1-6, DALL-E's own prompts are rewritten by ChatGPT before the underlying image model ever receives them — a real, useful convergence with prompt1's own conversational territory, but one where the actual text conditioning the diffusion model is never directly visible to the person iterating. A capstone built on DALL-E would be iterating on conversational instructions to ChatGPT, not directly on the descriptor- level prompt actually driving generation — a genuinely different, less mechanically transparent kind of iteration than this chapter's own five passes demonstrate. WHAT THE SCOPE NOTE ADMITS WOULD NEED TO CHANGE ------------------------------ Per this chapter's own honest scope note, "the underlying six-category vocabulary (imgai1-3) transfers, the exact syntax and available controls don't" — for Midjourney, Pass 3's negative prompt and CFG scale would become --no and --stylize instead, described behaviorally rather than via the exact formula this chapter used; for DALL-E, most of Passes 3-5 (negative prompt formula, term weighting syntax, checkpoint swapping) would have no direct equivalent at all, since none of those controls are exposed through ChatGPT's own conversational interface. WHY THIS MAKES STABLE DIFFUSION THE RIGHT CHOICE FOR THIS SPECIFIC CAPSTONE ------------------------------ Because this capstone's own explicit goal was to walk through a five-pass iteration workflow with each pass's mechanism explained precisely (not just described behaviorally), only Stable Diffusion — the one tool this course could describe mechanically throughout imgai1-5 — could support that level of explanatory precision at every single pass. The other two tools would have produced a real, working iteration workflow, but a shallower, less mechanically-grounded one. WHY THIS WORKS AS AN ANSWER ------------------------------ It uses the chapter's own stated reasoning for choosing Stable Diffusion, explains concretely what would have been lost by choosing Midjourney (behavioral-only description) or DALL-E (mediated, non- transparent prompting) instead, and ties this back to the scope note's own honest admission about what would need to change for another tool.